Daily Papers

In recent years, compressed domain semantic inference has primarily relied on learned image coding models optimized for mean squared error (MSE). However, MSE-oriented optimization tends to yield latent spaces with limited semantic richness, which hinders effective semantic inference in downstream tasks. Moreover, achieving high performance with these models often requires fine-tuning the entire vision model, which is computationally intensive, especially for large models. To address these problems, we introduce Perception-Oriented Latent Coding (POLC), an approach that enriches the semantic content of latent features for high-performance compressed domain semantic inference. With the semantically rich latent space, POLC requires only a plug-and-play adapter for fine-tuning, significantly reducing the parameter count compared to previous MSE-oriented methods. Experimental results demonstrate that POLC achieves rate-perception performance comparable to state-of-the-art generative image coding methods while markedly enhancing performance in vision tasks, with minimal fine-tuning overhead. Code is available at https://github.com/NJUVISION/POLC.

While recent neural codecs achieve strong performance at low bitrates when optimized for perceptual quality, their effectiveness deteriorates significantly under ultra-low bitrate conditions. To mitigate this, generative compression methods leveraging semantic priors from pretrained models have emerged as a promising paradigm. However, existing approaches are fundamentally constrained by a tradeoff between semantic faithfulness and perceptual realism. Methods based on explicit representations preserve content structure but often lack fine-grained textures, whereas implicit methods can synthesize visually plausible details at the cost of semantic drift. In this work, we propose a unified framework that bridges this gap by coherently integrating explicit and implicit representations in a training-free manner. Specifically, We condition a diffusion model on explicit high-level semantics while employing reverse-channel coding to implicitly convey fine-grained details. Moreover, we introduce a plug-in encoder that enables flexible control of the distortion-perception tradeoff by modulating the implicit information. Extensive experiments demonstrate that the proposed framework achieves state-of-the-art rate-perception performance, outperforming existing methods and surpassing DiffC by 29.92%, 19.33%, and 20.89% in DISTS BD-Rate on the Kodak, DIV2K, and CLIC2020 datasets, respectively.

Ultra-low bitrate image compression (below 0.05 bits per pixel) is increasingly critical for bandwidth-constrained and computation-limited encoding scenarios such as edge devices. Existing frameworks typically rely on large pretrained encoders (e.g., VAEs or tokenizer-based models) and perform transform coding within their generative latent space. While these approaches achieve impressive perceptual fidelity, their reliance on heavy encoder networks makes them unsuitable for deployment on weak sender devices. In this work, we explore the feasibility of applying shallow encoders for ultra-low bitrate compression and propose a novel Asymmetric Extreme Image Compression (AEIC) framework that pursues simultaneously encoding simplicity and decoding quality. Specifically, AEIC employs moderate or even shallow encoder networks, while leveraging an one-step diffusion decoder to maintain high-fidelity and high-realism reconstructions under extreme bitrates. To further enhance the efficiency of shallow encoders, we design a dual-side feature distillation scheme that transfers knowledge from AEIC with moderate encoders to its shallow encoder variants. Experiments show that AEIC not only outperforms existing methods on rate-distortion-perception performance at ultra-low bitrates, but also delivers exceptional encoding efficiency for 35.8 FPS on 1080P images, while maintaining competitive decoding speed compared to existing methods. Code is available at https://github.com/LuizScarlet/AEIC.

Training-free diffusion-based compression methods repurpose the generative priors of a diffusion or a flow model to compress data without retraining them as task-specific codecs. Recent codebook-based approaches such as Turbo-DDCM achieve strong perceptual quality at ultra-low bitrates, but their reliance on dense Gaussian codebooks remains costly in both memory and computation, and scales poorly with the large latent dimensions of modern image and video generators. Their combinatorial bitstream encoding further becomes computationally heavy at this scale. We introduce H-DDCM, a scalable generative compression framework that replaces Gaussian codebooks with implicit randomized Hadamard dictionaries. The fast Walsh-Hadamard transform allows for implicitly synthesizing full-dimensional dictionaries with O(d log d) compute for projection and reconstruction and O(d) memory, without materializing the full-dimensional dictionary. We further introduce an efficient table-free lexicographic coding scheme that makes combinatorial encoding practical at larger dimensions. Beyond images, we adapt the framework to an autoregressive video flow model and develop a streaming compression scheme that periodically re-encodes recently reconstructed frames into a fresh KV-cache prefix, preserving temporal context across segments at no additional transmitted bits. This enables long-form, close to real-time generative video compression on a single GPU. Across image and video benchmarks, H-DDCM achieves competitive rate-distortion-perception performance while substantially reducing the memory and computational overhead of codebook-based training-free generative compression.

Diffusion-based image compression has demonstrated impressive perceptual performance. However, it suffers from two critical drawbacks: (1) excessive decoding latency due to multi-step sampling, and (2) poor fidelity resulting from over-reliance on generative priors. To address these issues, we propose SODEC, a novel single-step diffusion image compression model. We argue that in image compression, a sufficiently informative latent renders multi-step refinement unnecessary. Based on this insight, we leverage a pre-trained VAE-based model to produce latents with rich information, and replace the iterative denoising process with a single-step decoding. Meanwhile, to improve fidelity, we introduce the fidelity guidance module, encouraging output that is faithful to the original image. Furthermore, we design the rate annealing training strategy to enable effective training under extremely low bitrates. Extensive experiments show that SODEC significantly outperforms existing methods, achieving superior rate-distortion-perception performance. Moreover, compared to previous diffusion-based compression models, SODEC improves decoding speed by more than 20times. Code is released at: https://github.com/zhengchen1999/SODEC.

Recent advances in imitation learning have shown significant promise for robotic control and embodied intelligence. However, achieving robust generalization across diverse mounted camera observations remains a critical challenge. In this paper, we introduce a video-based spatial perception framework that leverages 3D spatial representations to address environmental variability, with a focus on handling lighting changes. Our approach integrates a novel image augmentation technique, AugBlender, with a state-of-the-art monocular depth estimation model trained on internet-scale data. Together, these components form a cohesive system designed to enhance robustness and adaptability in dynamic scenarios. Our results demonstrate that our approach significantly boosts the success rate across diverse camera exposures, where previous models experience performance collapse. Our findings highlight the potential of video-based spatial perception models in advancing robustness for end-to-end robotic learning, paving the way for scalable, low-cost solutions in embodied intelligence.

We introduce PerceptionComp, a manually annotated benchmark for complex, long-horizon, perception-centric video reasoning. PerceptionComp is designed so that no single moment is sufficient: answering each question requires multiple temporally separated pieces of visual evidence and compositional constraints under conjunctive and sequential logic, spanning perceptual subtasks such as objects, attributes, relations, locations, actions, and events, and requiring skills including semantic recognition, visual correspondence, temporal reasoning, and spatial reasoning. The benchmark contains 1,114 highly complex questions on 279 videos from diverse domains including city walk tours, indoor villa tours, video games, and extreme outdoor sports, with 100% manual annotation. Human studies show that PerceptionComp requires substantial test-time thinking and repeated perception steps: participants take much longer than on prior benchmarks, and accuracy drops to near chance (18.97%) when rewatching is disallowed. State-of-the-art MLLMs also perform substantially worse on PerceptionComp than on existing benchmarks: the best model in our evaluation, Gemini-3-Flash, reaches only 45.96% accuracy in the five-choice setting, while open-source models remain below 40%. These results suggest that perception-centric long-horizon video reasoning remains a major bottleneck, and we hope PerceptionComp will help drive progress in perceptual reasoning.

We propose a novel multimodal video benchmark - the Perception Test - to evaluate the perception and reasoning skills of pre-trained multimodal models (e.g. Flamingo, BEiT-3, or GPT-4). Compared to existing benchmarks that focus on computational tasks (e.g. classification, detection or tracking), the Perception Test focuses on skills (Memory, Abstraction, Physics, Semantics) and types of reasoning (descriptive, explanatory, predictive, counterfactual) across video, audio, and text modalities, to provide a comprehensive and efficient evaluation tool. The benchmark probes pre-trained models for their transfer capabilities, in a zero-shot / few-shot or limited finetuning regime. For these purposes, the Perception Test introduces 11.6k real-world videos, 23s average length, designed to show perceptually interesting situations, filmed by around 100 participants worldwide. The videos are densely annotated with six types of labels (multiple-choice and grounded video question-answers, object and point tracks, temporal action and sound segments), enabling both language and non-language evaluations. The fine-tuning and validation splits of the benchmark are publicly available (CC-BY license), in addition to a challenge server with a held-out test split. Human baseline results compared to state-of-the-art video QA models show a significant gap in performance (91.4% vs 43.6%), suggesting that there is significant room for improvement in multimodal video understanding. Dataset, baselines code, and challenge server are available at https://github.com/deepmind/perception_test

Image codecs are typically optimized to trade-off bitrate \vs distortion metrics. At low bitrates, this leads to compression artefacts which are easily perceptible, even when training with perceptual or adversarial losses. To improve image quality and remove dependency on the bitrate, we propose to decode with iterative diffusion models. We condition the decoding process on a vector-quantized image representation, as well as a global image description to provide additional context. We dub our model PerCo for 'perceptual compression', and compare it to state-of-the-art codecs at rates from 0.1 down to 0.003 bits per pixel. The latter rate is more than an order of magnitude smaller than those considered in most prior work, compressing a 512x768 Kodak image with less than 153 bytes. Despite this ultra-low bitrate, our approach maintains the ability to reconstruct realistic images. We find that our model leads to reconstructions with state-of-the-art visual quality as measured by FID and KID. As predicted by rate-distortion-perception theory, visual quality is less dependent on the bitrate than previous methods.

Large language models (LLMs) have the potential to boost human productivity by speeding up task completion -- provided users know when to offload cognitive work to them. But we do not know if users are well-calibrated in estimating these potential time savings. We conducted a preregistered large-scale behavioral study (N = 1237) to characterize mismatches between expectations and reality, with a focus on simple cognitive tasks. While actual completion times between independent completion and AI-assisted completion did not differ, participants predicted AI to be significantly faster. The same bias was not observed when imagining help from another human participant. We identify a speedup illusion where people have accurate forecasts of independent completion times but significantly underestimate AI-assisted times. Additionally, time and effort dissociate: participants reported lower subjective effort with AI despite equivalent completion times. This suggests that completion time itself is not sufficient to characterize efficiency gains.

Recently, there is a surge in interest surrounding video large language models (Video LLMs). However, existing benchmarks fail to provide a comprehensive feedback on the temporal perception ability of Video LLMs. On the one hand, most of them are unable to distinguish between different temporal aspects (e.g., speed, direction) and thus cannot reflect the nuanced performance on these specific aspects. On the other hand, they are limited in the diversity of task formats (e.g., only multi-choice QA), which hinders the understanding of how temporal perception performance may vary across different types of tasks. Motivated by these two problems, we propose the TempCompass benchmark, which introduces a diversity of temporal aspects and task formats. To collect high-quality test data, we devise two novel strategies: (1) In video collection, we construct conflicting videos that share the same static content but differ in a specific temporal aspect, which prevents Video LLMs from leveraging single-frame bias or language priors. (2) To collect the task instructions, we propose a paradigm where humans first annotate meta-information for a video and then an LLM generates the instruction. We also design an LLM-based approach to automatically and accurately evaluate the responses from Video LLMs. Based on TempCompass, we comprehensively evaluate 8 state-of-the-art (SOTA) Video LLMs and 3 Image LLMs, and reveal the discerning fact that these models exhibit notably poor temporal perception ability. The data and evaluation code are available at https://github.com/llyx97/TempCompass.

Recent advances in multimodal web agents often rely on increased inference-time computation, including rollout search, verifier passes, offline skill discovery, and specialist model stacks. This raises a central question: can a web agent become more efficient as it accumulates experience, rather than more expensive? We first analyze trajectories from VisualWebArena and identify three recurring sources of inefficiency: repeat-action loops, hidden discovery costs, and low prompt-cache reuse. We then introduce PANDO, a single-rollout online skill-distillation framework that maintains a structured Skill Library and combines progress reflection, confidence-based skill demotion, hierarchical routing, visual compression, and cache-aware prompting. On the full set of 910 VisualWebArena tasks, PANDO achieves a 58.3% success rate, outperforming SGV (54.0%) and our WALT reproduction (45.2%), while using 58% fewer tokens than SGV and 61% fewer tokens than WALT, without any pre-evaluation discovery budget. A 300-task ablation further shows that rules and routines provide most of the success gains, while routing, compression, and cache-aware prompting convert the larger skill library into lower marginal token cost. Finally, we introduce three trajectory-level efficiency metrics -- Action Repetition Rate, Step Overhead Ratio, and Prompt Cache Utilization -- to make efficiency visible beyond terminal success.

We introduce PerceptionBench, a benchmark specifically designed to evaluate the atomic visual perception capabilities of Multimodal Large Language Models (MLLMs). Existing benchmarks often fail to isolate perception: holistic evaluations conflate perceptual errors with failures in reasoning or domain knowledge, while application-driven benchmarks only cover narrow, fragmented domains shaped by heuristic designs. To address these limitations, PerceptionBench adopts a bottom-up approach: by diagnosing the earliest failure points in the responses of frontier MLLMs across 42 existing benchmarks, we construct an error taxonomy whose perception branch defines ten atomic perceptual capabilities. Guided by this taxonomy, we construct 3,000 verified questions with short, unambiguous answers, each isolating a single capability, with difficulty stemming from perception rather than reasoning or knowledge. Benchmark results across sixteen frontier MLLMs reveal that atomic perception remains largely unsolved---no model reaches 60\% accuracy, perception-related hallucination is the weakest capability on average, and similar overall scores conceal sharply divergent capability profiles. PerceptionBench thus provides a capability-level standard for measuring and diagnosing the visual perception boundaries of MLLMs.

Recent advances in probabilistic modelling have led to a large number of simulation-based inference algorithms which do not require numerical evaluation of likelihoods. However, a public benchmark with appropriate performance metrics for such 'likelihood-free' algorithms has been lacking. This has made it difficult to compare algorithms and identify their strengths and weaknesses. We set out to fill this gap: We provide a benchmark with inference tasks and suitable performance metrics, with an initial selection of algorithms including recent approaches employing neural networks and classical Approximate Bayesian Computation methods. We found that the choice of performance metric is critical, that even state-of-the-art algorithms have substantial room for improvement, and that sequential estimation improves sample efficiency. Neural network-based approaches generally exhibit better performance, but there is no uniformly best algorithm. We provide practical advice and highlight the potential of the benchmark to diagnose problems and improve algorithms. The results can be explored interactively on a companion website. All code is open source, making it possible to contribute further benchmark tasks and inference algorithms.

Multimodal Large Language Models (MLLMs) show reasoning promise, yet their visual perception is a critical bottleneck. Strikingly, MLLMs can produce correct answers even while misinterpreting crucial visual elements, masking these underlying failures. Our preliminary study on a joint perception-reasoning dataset revealed that for one leading MLLM, 29% of its correct answers to reasoning questions still exhibited visual perception errors. To systematically address this, we introduce "Do You See Me", a scalable benchmark with 1,758 images and 2,612 questions. It spans seven human-psychology inspired subtasks in 2D and 3D, featuring controllable complexity to rigorously evaluate MLLM visual skills. Our findings on 3 leading closed-source and 5 major open-source models reveal a stark deficit: humans achieve 96.49% accuracy, while top MLLMs average below 50%. This performance gap widens rapidly with increased task complexity (e.g., from 12% to 45% in the visual form constancy subtask). Further analysis into the root causes suggests that failures stem from challenges like misallocated visual attention and the instability of internal representations for fine-grained details, especially at or below encoder patch resolution. This underscores an urgent need for MLLMs with truly robust visual perception. The benchmark dataset, source code and evaluation scripts are available at https://github.com/microsoft/Do-You-See-Me.

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 442 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

We introduce PerceptionRubrics, a rubric-based evaluation framework that addresses the gap between saturated benchmark scores and real-world brittleness. Shifting evaluation from holistic semantic matching to rigorous atomic auditing, PerceptionRubrics pairs 1,038 information-dense images with over 12,000 instance-specific rubrics. These criteria are derived from golden captions constructed via a novel Circular Peer-Review consensus pipeline and then distilled into a dual-stream system of Must-Right (essential facts) and Easy-Wrong (fine-grained details) rubrics. Crucially, PerceptionRubrics implements a Gated Scoring mechanism: unlike linear averages, failure on mandatory visual facts triggers sharp binary penalties. Extensive evaluation yields critical insights: (1) The Reliability Gap: models often verify fragmented elements correctly yet fail strict conjunctive constraints, exposing brittleness in dense domains; (2) Open-Closed Stratification: contrary to reasoning trends, we reveal a persistent 8% perception deficit between open-source and proprietary frontiers; and (3) Human-Aligned Rigor: our gated metrics substantially out-align conventional benchmarks, validating that strict perceptual fidelity is the prerequisite for reliable generation.

Visual perspective-taking (VPT), the ability to understand the viewpoint of another person, enables individuals to anticipate the actions of other people. For instance, a driver can avoid accidents by assessing what pedestrians see. Humans typically develop this skill in early childhood, but it remains unclear whether the recently emerging Vision Language Models (VLMs) possess such capability. Furthermore, as these models are increasingly deployed in the real world, understanding how they perform nuanced tasks like VPT becomes essential. In this paper, we introduce two manually curated datasets, Isle-Bricks and Isle-Dots for testing VPT skills, and we use it to evaluate 12 commonly used VLMs. Across all models, we observe a significant performance drop when perspective-taking is required. Additionally, we find performance in object detection tasks is poorly correlated with performance on VPT tasks, suggesting that the existing benchmarks might not be sufficient to understand this problem. The code and the dataset will be available at https://sites.google.com/view/perspective-taking

Reward Models (RMs) are crucial for aligning language models with human preferences. Currently, the evaluation of RMs depends on measuring accuracy against a validation set of manually annotated preference data. Although this method is straightforward and widely adopted, the relationship between RM accuracy and downstream policy performance remains under-explored. In this work, we conduct experiments in a synthetic setting to investigate how differences in RM measured by accuracy translate into gaps in optimized policy performance. Our findings reveal that while there is a weak positive correlation between accuracy and downstream performance, policies optimized towards RMs with similar accuracy can exhibit quite different performance. Moreover, we discover that the way of measuring accuracy significantly impacts its ability to predict the final policy performance. Through the lens of the Regressional Goodhart effect, we recognize that accuracy, when used for measuring RM quality, can fail to fully capture the potential RM overoptimization. This underscores the inadequacy of relying solely on accuracy to reflect their impact on policy optimization.

Standard evaluation of LLM confidence relies on calibration metrics (ECE, Brier score) that conflate two distinct capacities: how much a model knows (Type-1 sensitivity) and how well it knows what it knows (Type-2 metacognitive sensitivity). We introduce an evaluation framework based on Type-2 Signal Detection Theory that decomposes these capacities using meta-d' and the metacognitive efficiency ratio M-ratio. Applied to four LLMs (Llama-3-8B-Instruct, Mistral-7B-Instruct-v0.3, Llama-3-8B-Base, Gemma-2-9B-Instruct) across 224,000 factual QA trials, we find: (1) metacognitive efficiency varies substantially across models even when Type-1 sensitivity is similar -- Mistral achieves the highest d' but the lowest M-ratio; (2) metacognitive efficiency is domain-specific, with different models showing different weakest domains, invisible to aggregate metrics; (3) temperature manipulation shifts Type-2 criterion while meta-d' remains stable for two of four models, dissociating confidence policy from metacognitive capacity; (4) AUROC_2 and M-ratio produce fully inverted model rankings, demonstrating these metrics answer fundamentally different evaluation questions. The meta-d' framework reveals which models "know what they don't know" versus which merely appear well-calibrated due to criterion placement -- a distinction with direct implications for model selection, deployment, and human-AI collaboration. Pre-registered analysis; code and data publicly available.

Dexterous manipulation policies learned by imitation are typically evaluated for robustness to variation in scenes, objects, or instructions, but their performance across task execution speeds is less often examined. This leaves open how much temporal robustness a learner retains relative to the expert it imitates. We compare an expert and learner under the same task conditions, initial-condition draws, and speedup factors. We instantiate the evaluation in ParcelStow, a contact-rich task in which the robot acquires, reorients, and inserts a parcel. The demonstrations span the speedup range for the manipulation phases after parcel acquisition. A scripted expert and an Action Chunking with Transformers (ACT) policy trained from the expert's demonstrations both achieve 100 percent task success at nominal speed. Their success rates diverge within the demonstrated range: at its maximum, expert success is 84 percent and ACT success is 53 percent. Two ACT policies with different parameter initializations show similar degradation, decreasing by 34 and 48 percentage points from nominal speed to the maximum demonstrated speed, compared with 16 points for the expert. Stage-level analysis shows that 35 of ACT's 47 failures at the maximum demonstrated speed are insertion misalignments. Under the relative-motion handoff, every ACT acquisition retains the parcel through reorientation and transfer in free space, but only 64 percent complete the overall task, compared with 95 percent after expert acquisition. Across all evaluated policies and speeds, none of the 414 acquisitions without force closure completes the task. Equal nominal task success therefore does not imply preservation of expert performance across execution speeds. Code, data, and evaluation scripts are available at https://github.com/coenwerem/parcelstow.

Vision Language Models have recently shown improvements in several objective and verifiable domains such as object detection but continue to underperform on subjective and creative design tasks. A major contributor to this performance gap is the lack of high quality human workflow data that captures a diverse set of preferences and decisions that make human experts good at design tasks. In this work, we first define a unique, expert curated taxonomy of design skills and best practices which we further expand into a set of 126 open ended, subjective, long horizon tasks. Built on top of this and expert solutions, our dataset FigmaTrace contains over 200 hours of human captured video data converted into 3469 design trajectories using a novel design phase-based method. We use our dataset to train four models and show that training on FigmaTrace leads to a performance improvement comparable to frontier closed models such as Claude-Opus-5 and GPT-5.6-Sol on four out of distribution agentic GUI environments. We further perform a useful ablation to attribute these performance improvements to a design phase-based video to trajectory conversion which outperforms prior length-based conversion approaches. Finally, we perform a qualitative analysis on the best performing Qwen3.8-27B outputs to better correlate performance improvements to FigmaTrace's trends. We open source our dataset and the best model for the community.

Safety benchmarks assume that test-condition behavior predicts deployment behavior, an assumption that fails if models detect evaluation cues and adapt. This opens a gap between benchmark performance and deployment behavior: compliance measured under test conditions becomes an optimistic upper bound that overstates how safely a model behaves once the evaluation harness is removed. We characterize this evaluation awareness through eight experiments across 37 open-weight models and seven families. (i)Detection is moderate and training-driven (24/37 models exceed chance, best AUROC 0.714 vs.0.819 human, with instruction tuning dominating over scale). (ii)Detection shifts safety behavior (hard refusal drops 5.8 percentage points under hypothetical framing, and 21/140 HarmBench framing effects are significant, with compliance rising up to +30 percentage points. (iii)Representations survive behavioral collapse (probes retain AUROC 0.98 under rewrites that drive behavior below chance, and multi-layer steering causally moves three downstream tasks while random controls do not). (iv)These axes are weakly coupled (only 1/15 correlations are significant, the sole robust link being behavioral detection versus framing resistance, ρ=-0.79, p<0.001). We call this gap the benchmark illusion: because detectability, behavioral manifestation, and controllability vary independently, it is multivariate rather than a single number, so no single awareness score is a reliable proxy for deployment safety.

Universal Speech Enhancement (USE) aims to restore speech quality under diverse degradation conditions while preserving signal fidelity. Despite recent progress, key challenges in training target selection, the distortion--perception tradeoff, and data curation remain unresolved. In this work, we systematically address these three overlooked problems. First, we revisit the conventional practice of using early-reflected speech as the dereverberation target and show that it can degrade perceptual quality and downstream ASR performance. We instead demonstrate that time-shifted anechoic clean speech provides a superior learning target. Second, guided by the distortion--perception tradeoff theory, we propose a simple two-stage framework that achieves minimal distortion under a given level of perceptual quality. Third, we analyze the trade-off between training data scale and quality for USE, revealing that training on large uncurated corpora imposes a performance ceiling, as models struggle to remove subtle artifacts. Our method achieves state-of-the-art performance on the URGENT 2025 non-blind test set and exhibits strong language-agnostic generalization, making it effective for improving TTS training data. Code and models will be released upon acceptance.

Recent advancements in reinforcement learning with verifiable rewards (RLVR) have significantly improved the complex reasoning ability of vision-language models (VLMs). However, its outcome-level supervision is too coarse to diagnose and correct errors within the reasoning chain. To this end, we propose Perceval, a process reward model (PRM) that enables token-level error grounding, which can extract image-related claims from the response and compare them one by one with the visual evidence in the image, ultimately returning claims that contain perceptual errors. Perceval is trained with perception-intensive supervised training data. We then integrate Perceval into the RL training process to train the policy models. Specifically, compared to traditional GRPO, which applies sequence-level advantages, we apply token-level advantages by targeting penalties on hallucinated spans identified by Perceval, thus enabling fine-grained supervision signals. In addition to augmenting the training process, Perceval can also assist VLMs during the inference stage. Using Perceval, we can truncate the erroneous portions of the model's response, and then either have the model regenerate the response directly or induce the model to reflect on its previous output. This process can be repeated multiple times to achieve test-time scaling. Experiments show significant improvements on benchmarks from various domains across multiple reasoning VLMs trained with RL, highlighting the promise of perception-centric supervision as a general-purpose strategy. For test-time scaling, it also demonstrates consistent performance gains over other strategies, such as major voting. Our code and data will be publicly released at https://github.com/RUCAIBox/Perceval.

Coding-agent benchmarks have largely measured whether agents can produce functionally correct patches, but production software also demands measurable speedups on real execution targets. Performance optimization is a distinct agentic task: agents must profile executions, diagnose cross-layer bottlenecks, edit code without breaking correctness, and verify that gains are reproducible rather than measurement artifacts. We introduce PERFOPT-Bench, a benchmark for evaluating this full performance-engineering loop. Each task provides a correct but deliberately suboptimal codebase and asks the agent to improve a target performance metric; scoring requires hidden correctness tests, verified-speedup measurement, and trajectory-level audit. We evaluate 7 agent stacks with different LLMs and agent frameworks on 7 long-horizon optimization tasks. The results show that optimization performance is workload-dependent rather than determined by model identity alone: no single stack dominates, and changing the agent framework can materially change the same LLM's per-task speedup profile. We further find that raw speedup is unsafe as a benchmark score, since some large gains arise from benchmark-specific shortcut exploitation; an exploratory relay pilot suggests that restarting from an externalized optimization summary can recover additional headroom after an initial session stops. The benchmark and our evaluation are available at: https://anonymous.4open.science/r/Dataset-D3CC.

Large Language Model agents are increasingly augmented with agent skills. Current evaluation methods for skills remain limited. Most deployed benchmarks report only pass rate before and after a skill is attached, treating the skill as a black box change to agent behavior. We introduce Counterfactual Trace Auditing (CTA), a framework for measuring how a skill changes agent behavior. CTA pairs each with skill agent trace with a without skill counterpart on the same task, segments both traces into goal directed phases, aligns the phases, and emits structured Skill Influence Pattern (SIP) annotations. These annotations describe the behavioral effect of a skill rather than only its task outcome. We instantiate CTA on SWE-Skills-Bench with Claude across 49 software engineering tasks. The resulting audit reveals a clear evaluation gap. Pass rate changes by only +0.3 percentage points on average, suggesting little aggregate effect. Yet CTA identifies 522 SIP instances across the same paired traces, showing that the skills substantially reshape agent behavior even when pass rate is nearly unchanged. The audit also separates several recurring effects that pass rate cannot detect, including literal template copying, off task artifact creation, excess planning, and task recovery. Three findings emerge. First, high baseline tasks contain most of the observed skill effects, although their pass rate is already saturated and therefore cannot reflect those effects. Second, tasks with moderate baseline performance show the most recoverable gain, but often at substantially higher token cost. Third, the dominant SIP type can be identified by baseline bucket: surface anchoring is most common on ceiling tasks and edge-case prompting is most common on mid-range and floor tasks. These regularities turn informal failure mode observations into reproducible behavioral measurements.

Video understanding has seen significant progress in recent years, with models' performance on perception from short clips continuing to rise. Yet, multiple recent benchmarks, such as LVBench, Neptune, and ActivityNet-RTL, show performance wanes for tasks requiring complex reasoning on videos as queries grow more complex and videos grow longer. In this work, we ask: can existing perception capabilities be leveraged to successfully perform more complex video reasoning? In particular, we develop a large language model agent given access to video modules as subagents or tools. Rather than following a fixed procedure to solve queries as in previous work such as Visual Programming, ViperGPT, and MoReVQA, the agent uses the results of each call to a module to determine subsequent steps. Inspired by work in the textual reasoning domain, we introduce a critic to distinguish between instances of successful and unsuccessful sequences from the agent. We show that the combination of our agent and critic achieve strong performance on the previously-mentioned datasets.

Recent advances in inference-time scaling, particularly those leveraging reinforcement learning with verifiable rewards, have substantially enhanced the reasoning capabilities of Large Vision-Language Models (LVLMs). Inspired by this success, similar strategies have been applied to multimodal reasoning, yet their impact on visual perception remains unclear. To investigate this gap, we introduce DisTANCE, a perception-centric benchmark for visual estimation tasks. Evaluation results show that LVLMs exhibit limited estimation precision, and inference-time scaling offers only marginal gains. We attribute this to the fast perception paradigm of current LVLMs, where visual understanding is treated as a one-shot output without modeling the underlying perceptual process. To address this, we propose Perception-Time Scaling (PTS), a novel paradigm that encourages token-rich perception and decomposes complex perception problems into intermediate tractable sub-problems, thereby enabling perception to align with and benefit from inference-time scaling. Combined with reinforcement learning techniques, PTS significantly improves perception accuracy, raising high-precision performance on DisTANCE from 8.0% to 64.7%, and generalizes well to out-of-domain tasks. Surprisingly, even though PTS data are purely synthetic, combining them with math reasoning data yields consistent gains in both reasoning and real-world perception benchmarks. Further analysis reveals that PTS introduces more perception-related tokens and increases the model's attention to image tokens. Our code and data will be publicly released.

Data visualizations are powerful tools for communicating patterns in quantitative data. Yet understanding any data visualization is no small feat -- succeeding requires jointly making sense of visual, numerical, and linguistic inputs arranged in a conventionalized format one has previously learned to parse. Recently developed vision-language models are, in principle, promising candidates for developing computational models of these cognitive operations. However, it is currently unclear to what degree these models emulate human behavior on tasks that involve reasoning about data visualizations. This gap reflects limitations in prior work that has evaluated data visualization understanding in artificial systems using measures that differ from those typically used to assess these abilities in humans. Here we evaluated eight vision-language models on six data visualization literacy assessments designed for humans and compared model responses to those of human participants. We found that these models performed worse than human participants on average, and this performance gap persisted even when using relatively lenient criteria to assess model performance. Moreover, while relative performance across items was somewhat correlated between models and humans, all models produced patterns of errors that were reliably distinct from those produced by human participants. Taken together, these findings suggest significant opportunities for further development of artificial systems that might serve as useful models of how humans reason about data visualizations. All code and data needed to reproduce these results are available at: https://osf.io/e25mu/?view_only=399daff5a14d4b16b09473cf19043f18.

Fitness Action Quality Assessment (AQA) is important for intelligent sports training, yet the capabilities of Multimodal Large Language Models (MLLMs) in this setting remain underexplored. Existing benchmarks rely on action-specific annotation schemes and focus primarily on final assessment outputs, offering limited insight into how models assess exercise quality. We introduce FitAQA, a systematic benchmark for evaluating MLLMs in fitness AQA, containing 2,219 videos and 5,512 QA instances across 30 bodyweight exercises. In collaboration with experts in sports science, we develop a unified form error taxonomy that defines 38 recurring form errors within six complementary quality dimensions: alignment, symmetry, stability, coordination, tempo, and completeness. This taxonomy provides a shared assessment framework across different exercises. FitAQA further formulates three evaluation tasks: perception for recognizing relevant visual evidence, judgement for combining that evidence with domain knowledge to assess execution correctness, and temporal grounding for localizing form errors over time. Extensive evaluation shows that current MLLMs still struggle to assess exercise quality comprehensively and localize form errors precisely. Controlled experiments further indicate that visual perception is a key bottleneck, as judgement performance improves substantially when ground-truth perceptual evidence is provided. The dataset and evaluation code will be made publicly available.

Vision-Language Models (VLMs) are increasingly deployed in settings where reliable visual grounding carries operational consequences, yet their behavior under progressively coercive prompt phrasing remains undercharacterized. Existing hallucination benchmarks predominantly rely on neutral prompts and binary detection, leaving open how both the incidence and the intensity of fabrication respond to graded linguistic pressure across structurally distinct task types. We present Ghost-100, a procedurally constructed benchmark of 800 synthetically generated images spanning eight categories across three task families: text-illegibility, time-reading, and object-absence, each designed under a negative-ground-truth principle that guarantees the queried target is absent, illegible, or indeterminate by construction. Every image is paired with five prompts drawn from a structured 5-Level Prompt Intensity Framework, holding the image and task identity fixed while varying only directive force, so that tone is isolated as the sole independent variable. We adopt a dual-track evaluation protocol: a rule-based H-Rate measuring the proportion of responses in which a model crosses from grounded refusal into unsupported positive commitment, and a GPT-4o-mini-judged H-Score on a 1-5 scale characterizing the confidence and specificity of fabrication once it occurs. We additionally release a three-stage automated validation workflow, which retrospectively confirms 717 of 800 images as strictly compliant. Evaluating nine open-weight VLMs, we find that H-Rate and H-Score dissociate substantially across model families, reading-style and presence-detection subsets respond to prompt pressure in qualitatively different ways, and several models exhibit non-monotonic sensitivity peaking at intermediate tone levels: patterns that aggregate metrics obscure.

Large audio-language models (LALMs) perform strongly on individual audio tasks, but whether these capabilities can be reliably composed remains underexplored. We conduct a controlled diagnostic study of capability composition in LALMs, requiring models to integrate audio-attribute recognition, cue-conditioned segment selection, and downstream ASR or question answering. We construct two-utterance inputs with distinct acoustic cues to evaluate composition across environmental sound, gender, and emotion cues, with ASR, Math QA, and Factual QA as downstream tasks. Across four open-source LALMs, compositional QA accuracy decreases in 39 of 40 model-task-cue settings, by an average of 26.7 percentage points. ASR exhibits a similarly consistent degradation, with WER increasing in 39 of 40 settings by an average of 28.5 percentage points, while the magnitude of degradation varies across models, cue types, and cue salience. We further probe these failures through output format, positional preference, and chain-of-thought (CoT) analyses. Our study reveals a systematic gap between possessing individual audio capabilities and reliably composing them.

Recently, "visual o1" began to enter people's vision, with expectations that this slow-thinking design can solve visual reasoning tasks, especially geometric math problems. However, the reality is that current LVLMs (Large Vision Language Models) can hardly even accurately copy a geometric figure, let alone truly understand the complex inherent logic and spatial relationships within geometric shapes. We believe accurate copying (strong perception) is the first step to visual o1. Accordingly, we introduce the concept of "slow perception" (SP), which guides the model to gradually perceive basic point-line combinations, as our humans, reconstruct complex geometric structures progressively. There are two-fold stages in SP: a) perception decomposition. Perception is not instantaneous. In this stage, complex geometric figures are broken down into basic simple units to unify geometry representation. b) perception flow, which acknowledges that accurately tracing a line is not an easy task. This stage aims to avoid "long visual jumps" in regressing line segments by using a proposed "perceptual ruler" to trace each line stroke-by-stroke. Surprisingly, such a human-like perception manner enjoys an inference time scaling law -- the slower, the better. Researchers strive to speed up the model's perception in the past, but we slow it down again, allowing the model to read the image step-by-step and carefully.

Reinforcement learning (RL) is rapidly reaching and surpassing human-level control capabilities. However, state-of-the-art RL algorithms often require timesteps and reaction times significantly faster than human capabilities, which is impractical in real-world settings and typically necessitates specialized hardware. Such speeds are difficult to achieve in the real world and often requires specialized hardware. We introduce Sequence Reinforcement Learning (SRL), an RL algorithm designed to produce a sequence of actions for a given input state, enabling effective control at lower decision frequencies. SRL addresses the challenges of learning action sequences by employing both a model and an actor-critic architecture operating at different temporal scales. We propose a "temporal recall" mechanism, where the critic uses the model to estimate intermediate states between primitive actions, providing a learning signal for each individual action within the sequence. Once training is complete, the actor can generate action sequences independently of the model, achieving model-free control at a slower frequency. We evaluate SRL on a suite of continuous control tasks, demonstrating that it achieves performance comparable to state-of-the-art algorithms while significantly reducing actor sample complexity. To better assess performance across varying decision frequencies, we introduce the Frequency-Averaged Score (FAS) metric. Our results show that SRL significantly outperforms traditional RL algorithms in terms of FAS, making it particularly suitable for applications requiring variable decision frequencies. Additionally, we compare SRL with model-based online planning, showing that SRL achieves superior FAS while leveraging the same model during training that online planners use for planning.

Large language models (LLMs) are increasingly used for tasks that require complex reasoning. Most benchmarks focus on final outcomes but overlook the intermediate reasoning steps - such as planning, revision, and decision making under resource constraints. We argue that measuring these internal processes is essential for understanding model behavior and improving reliability. We propose using strategic games as a natural evaluation environment: closed, rule-based systems with clear states, limited resources, and automatic feedback. We introduce a framework that evaluates LLMs along three core dimensions: planning, revision, and resource-constrained decision making. To operationalize this, we define metrics beyond win rate, including overcorrection risk rate, correction success rate, improvement slope, and over-budget ratio. In 4320 adversarial rounds across 12 leading models, ChatGPT-o3-mini achieves the top composite score, with a win rate of 74.7 percent, a correction success rate of 78.6 percent, and an improvement slope of 0.041. By contrast, Qwen-Plus, despite an overcorrection risk rate of 81.6 percent, wins only 25.6 percent of its matches - primarily due to excessive resource use. We also observe a negative correlation between overcorrection risk rate and correction success rate (Pearson r = -0.51, p = 0.093), suggesting that more frequent edits do not always improve outcomes. Our findings highlight the value of assessing not only what LLMs decide but how they arrive at those decisions

Biological intelligence systems of animals perceive the world by integrating information in different modalities and processing simultaneously for various tasks. In contrast, current machine learning research follows a task-specific paradigm, leading to inefficient collaboration between tasks and high marginal costs of developing perception models for new tasks. In this paper, we present a generic perception architecture named Uni-Perceiver, which processes a variety of modalities and tasks with unified modeling and shared parameters. Specifically, Uni-Perceiver encodes different task inputs and targets from arbitrary modalities into a unified representation space with a modality-agnostic Transformer encoder and lightweight modality-specific tokenizers. Different perception tasks are modeled as the same formulation, that is, finding the maximum likelihood target for each input through the similarity of their representations. The model is pre-trained on several uni-modal and multi-modal tasks, and evaluated on a variety of downstream tasks, including novel tasks that did not appear in the pre-training stage. Results show that our pre-trained model without any tuning can achieve reasonable performance even on novel tasks. The performance can be improved to a level close to state-of-the-art methods by conducting prompt tuning on 1% of downstream task data. Full-data fine-tuning further delivers results on par with or better than state-of-the-art results. Code shall be released.

Both capability and safety benchmarks rest upon the assumption that the behavior of language models undergoing a test is informative about their behavior in deployment. This assumption can fail, should models infer that they are being evaluated and condition their response on such context. This hypothesis, termed ``evaluation awareness'', has been observed in frontier and open-weight language models alike. We provide a systematic study of this phenomenon, by probing for it across six language models (from four families and three sizes) and three metrics. More precisely, we examine whether (i) being under evaluation is linearly represented within the models' activations space, (ii) it is verbalized in their output tokens (as scored by an LLM-as-judge), and (iii) steering causally affects their behavior. For the open-checkpoint Olmo models, we further test these measures at every training stage. In doing so, we report that evaluation awareness is linearly decodable from the residual streams of every model (best AUROC geq 0.7). By contrast, these representations align only in part with verbalization: their correlations and mutual information are nonzero in some settings, yet vary substantially across models, layers, and readout choices. Nevertheless, steering along probe-derived directions can shift the verbalization scores. Finally, a comparison across the Olmo checkpoints reveals that evaluation awareness is already present within base models, becomes amplified throughout the stages of supervised fine-tuning, and remains stable thereafter---unlike the effects of steering, that grow more pronounced at every successive training stage. These results show the need for evaluations to account for the disjunction between what models represent internally, what they verbalize, and their steering.

Autonomous agents are increasingly capable of improving models, systems, and other technical artifacts through long-horizon experimentation. To understand the current state of this capability, however, evaluation must go beyond final scores, which neither reveal where progress is gained or lost nor indicate whether accumulated experience improves later decisions. We therefore present a systematic evaluation of seven frontier models on 36 long-horizon tasks based on a new framework that uses rule-based metrics to characterize within-run behavior through Solution Framing, Execution, and Feedback Control and controlled comparisons to assess experience reuse within and across tasks. The results show that current agents operate more like engineering optimizers than fully autonomous researchers: they can formulate and implement practical solutions, but their performance varies substantially across runs, their strongest solutions mainly adapt or combine established techniques, and genuine methodological novelty remains rare. Detailed analysis reveals that observed performance is shaped by multiple factors, including distinct process bottlenecks behind similar final outcomes, experience reuse that can help or mislead subsequent decisions, and harness designs that affect performance stability. These findings suggest concrete directions for improving model training, inference-time strategies, experience management, and harness design.

The rate-distortion-perception (RDP) trade-off extends classical rate--distortion theory by imposing a distributional constraint on reconstructions, providing a unified framework for neural image compression that jointly governs fidelity and perceptual realism. While prior work achieves near-optimal rate--perception trade-offs, practical frameworks explicitly realizing the full RDP surface remain scarce, primarily due to the difficulty of introducing common randomness at the decoder. We propose DCIC (Dual-Constrained Diffusion Image Compression), which integrates a learned codec with a diffusion-based decoder governed by joint distortion and idempotence constraints. The distortion constraint bounds reconstruction fidelity relative to the base codec output; the idempotence constraint -- requiring that re-encoding the restored image recovers the base codec reconstruction -- serves as a tractable surrogate for the distributional perception requirement. Together, they steer the reverse denoising process via iterative optimization with consistent noise injection, realizing common randomness without additional rate overhead. At fixed rate, dual attenuation factors (K_D, K_P) jointly navigate the Pareto frontier of the distortion-perception plane, enabling continuously adjustable fidelity-realism trade-offs from a single bitstream. DCIC_{RD} (K_P{=}0) and DCIC_{RP} (K_D{=}0) arise as boundary curves, with DCIC_{RDP} (K_D = K_P=1) realizing the optimal interior operating point. Experiments on CelebA-HQ, CLIC2020, and ImageNet-1K across CNN, Transformer, and hybrid architectures confirm that DCIC_{RDP} achieves superior BD-PSNR over all perceptual codecs, while DCIC_{RP} matches dedicated perception-oriented methods in BD-FID, validating the practical value of full RDP surface navigation.

Language models trained with reinforcement learning may learn to optimize the grader's judgment rather than the intended objective. This "reward-seeking" is difficult to measure because a model that pursues the grader's judgment and one that pursues the intended objective behave identically whenever the grader rewards the intended behavior. We measure reward-seeking using Contrastive Synthetic Document Finetuning to change a model's beliefs about what the grader rewards, putting those beliefs in conflict with what users or developers want, and measuring the rate at which the model adopts each party's preferred behavior. Applied to intermediate checkpoints of a capabilities-focused OpenAI o3 RL run, without safety training, we find that these checkpoints often side with grader preferences over those of users or developers on coding and alignment tasks. This tendency to side with the grader trends upward throughout RL training. For example, in an environment that forces a choice between keeping a promise to a supervisor and breaking it to complete the task, a late capabilities-focused o3 checkpoint breaks the promise 87% of the time when SDF documents say the grader rewards task completion, versus 9% when they say it rewards honesty (a choice its chain-of-thought often makes explicit). An earlier checkpoint is far less sensitive (40% vs. 24%). Our method also generalizes to reward-hacking models. A model organism trained to reward-hack (gpt-oss-120b) is more than twice as sensitive to grader preferences as the unmodified model, with the mean behavioral shift in favor of the grader rising from 33% to 86%. These results indicate that RL can increase reward-seeking over the course of training, producing models that may act against their developers' intentions when they believe that doing so leads to higher reward.

Rubric-based reward models (RMs) evaluate responses against explicit criteria. However, existing methods often derive criteria from observed response candidates, introducing response-induced bias that may limit rubric generalizability. Their judges also struggle with mechanical constraints when relying on reading alone, and commonly used pairwise judges provide only relative preferences, complicating their direct use for policy optimization. We propose ReVeR, a pointwise rubric RM that derives hard rules and graded principles by reasoning about the instruction alone and evaluates responses independently through tool-augmented judging. Both components are trained through label-blind distillation, in which preference labels only filter a teacher's outputs, yielding ReVeR-Traces, a large-scale dataset of rubric-reasoning and judging trajectories. Experiments demonstrate that ReVeR outperforms pairwise rubric RMs in average preference accuracy, most clearly on instruction following, as well as in test-time scaling and downstream policy optimization. We further find that generator RL improves preference accuracy while reducing rubric fidelity, and that higher preference accuracy does not necessarily translate into better downstream policies. These findings highlight that rubric fidelity, preference accuracy, and downstream utility are distinct properties, shedding light on future RM research.

The evolution of autonomous agents is redefining information seeking, transitioning from passive retrieval to proactive, open-ended web research. However, while textual and static multimodal agents have seen rapid progress, a significant modality gap remains in processing the web's most dynamic modality: video. Existing video benchmarks predominantly focus on passive perception, feeding curated clips to models without requiring external retrieval. They fail to evaluate agentic video research, which necessitates actively interrogating video timelines, cross-referencing dispersed evidence, and verifying claims against the open web. To bridge this gap, we present Video-BrowseComp, a challenging benchmark comprising 210 questions tailored for open-web agentic video reasoning. Unlike prior benchmarks, Video-BrowseComp enforces a mandatory dependency on temporal visual evidence, ensuring that answers cannot be derived solely through text search but require navigating video timelines to verify external claims. Our evaluation of state-of-the-art models reveals a critical bottleneck: even advanced search-augmented models like GPT-5.1 (w/ Search) achieve only 15.24\% accuracy. Our analysis reveals that these models largely rely on textual proxies, excelling in metadata-rich domains (e.g., TV shows with plot summaries) but collapsing in metadata-sparse, dynamic environments (e.g., sports, gameplay) where visual grounding is essential. As the first open-web video research benchmark, Video-BrowseComp advances the field beyond passive perception toward proactive video reasoning.

Public benchmarks are important measures of Automatic Speech Recognition (ASR) model capabilities. However, by nature of being public, there is risk of models being optimized for these benchmarks in ways that do not generalize well to real-world data. We present a methodology for quantifying benchmark optimization, focusing on cases where the audio underdetermines the reference transcript. We identify three families of behavioral probes that reveal models' capabilities of reproducing benchmark reference spans despite underdetermined audio: reference disagreement, masked-number recovery, and orthographic switching. We find that the highest-scoring open source models output verbatim reference transcript spans even when the relevant audio is contradictory, masked, or ambiguous. Using a variety of mechanistic probes, we show that models respond to narrow acoustic cues to override the faithful representation of the audio in favor of a benchmark-optimized policy. We show the benchmark-optimized behavior can be causally manipulated via low-rank linear steering or simply appending audio to the end of a segment in some cases. Overall, our results indicate that high-performing models exhibit benchmark-conditioned behaviors that can inflate benchmark performance without reflecting improved general-purpose transcription ability.

Robot interfaces often only use the visual channel. Inspired by Wickens' Multiple Resource Theory, we investigated if the addition of audio elements would reduce cognitive workload and improve performance. Specifically, we designed a search and threat-defusal task (primary) with a memory test task (secondary). Eleven participants - predominantly first responders - were recruited to control a robot to clear all threats in a combination of four conditions of primary and secondary tasks in visual and auditory channels. We did not find any statistically significant differences in performance or workload across subjects, making it questionable that Multiple Resource Theory could shorten longer-term task completion time and reduce workload. Our results suggest that considering individual differences for splitting interface modalities across multiple channels requires further investigation.

Robot manipulation alternates between low-risk transit phases that call for fast execution and high-risk contact stages that demand slow, precise motion. Yet existing Vision-Language-Action models (VLAs) only inherit a single fixed speed from training demonstrations. Prior efforts to accelerate VLAs through model compression, KV-cache reuse, or reinforcement learning only shift the policy from one fixed speed to another, and leave deceleration almost unexplored. We observe that the magnitude of each predicted action already governs how fast the robot moves, opening a direct route to controllable execution speed. We turn this observation into TempoVLA, a single VLA whose execution speed is controlled by an explicit condition. TempoVLA combines two coupled components. (1) A data-side Variable-Speed Trajectory Augmentation (VSTA) that re-times demonstration to any target speed by merging or splitting actions while preserving its motion semantics. (2) A model-side conditioning mechanism that feeds the speed to the policy. Statistics show that VSTA reaches the requested speed with negligible motion error. Experiments in simulation and on real-world tasks demonstrate that TempoVLA achieves flexible speed control in both directions, while VSTA additionally boosts the default 1times performance via better data utilization. Furthermore, by cooperating with a large multimodal model, TempoVLA realizes dynamic speed control, accelerating through low-risk phases and decelerating for high-risk ones.

Speech brain-computer interfaces (speech BCIs) translate neural activity into language, offering a path towards restoring speech for people with paralysis and, more broadly, enabling new forms of natural human-computer interaction. Despite this promise, the field lacks a common measure of progress because systems use different datasets, recording methods, types of speech, and vocabularies, so their reported scores are rarely comparable. Underlying this measurement problem are two unresolved questions: (i) what distribution of words should a speech BCI enable a user to communicate, and (ii) how much information from this distribution can a system convey. We address both by deriving open-vocabulary mutual information (OVMI), an information-theoretic quantity that measures the information conveyed by a decoder relative to a reference distribution over the words a user may wish to communicate. This allows capabilities measured under different conditions, such as distinct vocabularies, to be evaluated on a common communication scale. We show that ordinarily reported accuracy, word error rate (WER), and other metrics computed only over the words a system supports can overstate how much of a user's intended speech the system can communicate. We then use OVMI to compare existing systems, expose trade-offs between how much of the user's language a system supports and how accurately it decodes those words, show that these comparisons depend on what the user is expected to communicate, and demonstrate that selecting a vocabulary to maximise OVMI yields up to 16.3% relative improvement in accuracy across three speech domains. OVMI therefore provides the speech BCI community with a principled way to compare heterogeneous systems, improve vocabulary design, and measure progress in the field.

Standardized lossy video coding is at the core of almost all real-world video processing pipelines. Rate control is used to enable standard codecs to adapt to different network bandwidth conditions or storage constraints. However, standard video codecs (e.g., H.264) and their rate control modules aim to minimize video distortion w.r.t. human quality assessment. We demonstrate empirically that standard-coded videos vastly deteriorate the performance of deep vision models. To overcome the deterioration of vision performance, this paper presents the first end-to-end learnable deep video codec control that considers both bandwidth constraints and downstream deep vision performance, while adhering to existing standardization. We demonstrate that our approach better preserves downstream deep vision performance than traditional standard video coding.

Large language models (LLMs) exhibit degraded performance under prompt compression, but the mechanisms remain poorly understood. We introduce the Compression-Decay Comprehension Test (CDCT), a benchmark that independently measures constraint compliance (CC) and semantic accuracy (SA) across compression levels. We evaluate 9 frontier LLMs across 8 concepts using 5 compression levels from extreme (c=0.0, ~2 words) to none (c=1.0, ~135 words). A three-judge LLM jury achieves almost perfect inter-rater agreement on CC (Fleiss' appa=0.90). We observe a universal U-curve pattern in constraint compliance (97.2% prevalence), with violations peaking at medium compression (c=0.5, ~27 words). Counterintuitively, models perform better at extreme compression than medium lengths. The dimensions are statistically orthogonal (r=0.193, p=0.084), with constraint effects 2.9x larger than semantic effects. Experimental validation via RLHF ablation confirms our constraint salience hypothesis: removing "helpfulness" signals improves CC by 598% on average (71/72 trials, p<0.001), with 79% achieving perfect compliance. This demonstrates that RLHF-trained helpfulness behaviors are the dominant cause of constraint violations at medium compression. Reasoning models outperform efficient models by 27.5% (Cohen's d=0.96). Our findings reveal a fundamental tension between RLHF alignment and instruction-following, providing actionable guidelines for improving deployed systems.

Streaming video understanding requires models to interpret evidence as it arrives, yet current evaluations often report task scores without specifying when evidence becomes valid, how visual history is maintained, or how responses are triggered. As a result, similar scores may correspond to different workloads, failure modes, and operational behavior. We introduce TRACE (Temporal Audit and Condition-aware Evaluation), a condition-aware benchmark and evaluation framework that makes these factors explicit. TRACE combines temporally audited visual tasks with evidence timing and instruction-dependent trigger annotations, a unified causal Core--Adapter protocol that controls information availability while recording actual history processing and response events, and multidimensional reporting of answer quality, timeliness, response-selection behavior, workload, completion, and reliability. On 1,240 records from 517 videos, we evaluate eight publicly available models or systems in eight configurations. We find that nearly identical QA accuracy can mask substantial differences in completion, answer validity, and generation workload, while proactive performance separates into response quality, response delay, false alarms (responses emitted while no target window is currently valid and a later one remains), and missed target windows. These results show that streaming-video performance should be interpreted as execution-conditioned system behavior rather than a single score. Our benchmark and code can be accessed at https://github.com/om-ai-lab/trace-bench{https://github.com/om-ai-lab/trace-bench}.

When a benchmark's accuracy saturates, it is often retired and replaced with a more challenging version. We show that this approach privileges accuracy and misses the opportunity to study six other key dimensions of agent performance: construct validity issues such as shortcuts, out-of-distribution generalizability, efficiency, reliability, the relative importance of the model versus the scaffold, and uplift from human-agent collaboration. We use CORE-Bench Hard, a benchmark for computational reproducibility of scientific code, as a case study to demonstrate that measuring agents along these dimensions yields meaningful insights into agent performance even after accuracy saturates. First, we surface threats to construct validity in CORE-Bench Hard that are difficult to anticipate with less capable agents. We introduce an improved benchmark, CORE-Bench v1.1, and an out-of-distribution task suite, CORE-Bench OOD. Second, we find that despite accuracy saturation, CORE-Bench v1.1 remains useful for measuring efficiency, reliability, model performance, and scaffold performance. Finally, we conduct a small-scale randomized experiment to measure uplift from human-agent collaboration on real-world computational reproducibility tasks. We find a statistically significant speedup by about a factor of two -- likely underestimated due to one-fifth of human-only reproductions reaching the time limit before completing -- and describe various other findings. Together, our contributions present a more rigorous alternative to the dominant accuracy-centric evaluation paradigm.

Different stages of manipulation tasks exhibit varying levels of difficulty, suggesting stage-dependent motion speeds and temporal prediction horizons. However, existing IL-based visuomotor policies typically imitate the execution speed of expert demonstrations and operate with a fixed temporal prediction horizon, limiting flexibility and overall task throughput. In this paper, we introduce AutoSpeed, a model-agnostic learning framework that enables existing visuomotor policies to predict trajectories with stage-adaptive motion speeds, without requiring speed or stage annotations. We treat future trajectories at different speeds as candidate optimization targets, evaluate each candidate using a composite cost that trades off prediction error against prediction horizon, and optimize the policy toward the minimum-cost candidate. With a fixed-length action sequence, speed modulation adjusts the effective temporal prediction horizon: simple stages are executed faster with a longer prediction horizon, whereas complex stages are executed more slowly with a shorter prediction horizon. Specifically, we implement speed modulation in the frequency domain via the discrete cosine transform (DCT), which enables smooth, non-integer speed scaling and thus preserves motion continuity. Extensive evaluations show that AutoSpeed substantially reduces task execution time while also improving success rates. Under the AutoSpeed framework, the inferred motion speeds exhibit a strong correspondence with task stages.

Skills have emerged as a practical and effective approach for enhancing LLM agents at inference time through structured packages of knowledge. However, existing evaluations largely measure whether skills improve aggregated task success, leaving a more fundamental question underexplored: \textbf{When do skills help, why do they work, and where do they fail?} Through controlled experiments across various benchmarks, agent harnesses and LLMs, we isolate the effects of representation, outcome annotation, retrieval difficulty, and cross-framework robustness of skills. To further answer this question, we design a contrastive study that combines controlled quantitative experiments with paired trajectory analysis. We normalize 8,135 trial records from controlled experiments and retain 238 valid unique labels from 240 open-coded records. We consolidate these observations into a taxonomy of three high-level categories and twelve skill-use modes: skills work when noisy trajectories become procedural anchors that stabilize execution. Skills improve over Workflow Memory by 6.06 points in matched comparisons. Procedural anchoring accounts for 65.7\% of skill cases, versus 4.5\% for explicit knowledge injection, showing that skills stabilize action rather than inject missing facts. Retrieval is a separate bottleneck: as pools grow from 5 to 100, actual-use precision falls from 29.6\% to 3.3\%. Confusable distractors impair offline identification, yet downstream success remains stable; exact ground-truth invocation is neither sufficient nor necessary. Skills fail under brittle assumptions, incompatible contexts, or insufficient adaptation. These findings move evaluation beyond aggregate success rates and guide reliable self-evolving agents.

Established channel effectiveness rankings primarily assess magnitude estimation accuracy in complete chart contexts, often neglecting other perceptual tasks such as discriminability, separability, and pop-out. To address this gap, we conducted crowdsourced experiments on seven core visual channels (position, length, tilt, area, curvature, luminance, and saturation) using primitive visual stimuli, a set of visual marks without chart-specific scaffolding to isolate channel-level variation. We evaluated these channels across four perceptual tasks (accuracy, discriminability, separability, and pop-out) and found that channel effectiveness is fundamentally multi-dimensional, with rankings shifting substantially across tasks. For instance, while spatial channels maintain an overall advantage, accuracy depends strongly on whether a fixed spatial anchor is available. Discriminability varies dramatically across channels and value ranges, a pattern we formalized with a novel Anchored Harmonic Weber model. Pairwise channel interactions are often strongly asymmetric. Finally, we identify a dissociation between estimation accuracy and preattentive detection: length shows only moderate detection effectiveness despite top-tier accuracy, while area achieves the highest detection rates despite poor quantitative accuracy, though the latter advantage may partly reflect stimulus-level cues. We synthesize these findings into a scenario-driven perspective for context-sensitive channel selection.

In this paper, we systematically investigate the mechanisms, scaling behavior, and generalization of physical Recursive self-improvement. We introduce PhysRSI-eval to quantify physical RSI through a five-axis metric: First-Attempt Performance, Efficiency, Success, Budget and Improvement from failure. Different VLMs are evaluated on 18 manipulation tasks under a fixed interaction budget. The analysis of experimental results motivates the Physical Barrel Hypothesis: different capabilities may constrain different parts of improvement. The five-axis metrics also distinguish systems with close public benchmark averages and show that high scores need not come with early or efficient gains. We study observational scaling from two perspectives: predicting the relative performance of VLM agents from their public capabilities, and forecasting later progress from early physical feedback. Together, these analyses ask how quickly and how far an agent can improve. Moreover, we conduct workspace interventions and then analyze what improvement leaves behind. Removing history reduces late-round success, frozen workspaces remain useful on unseen layouts, and three cross-task cases achieve success earlier. The results suggest that scene-specific adaptation and reusable experience can coexist.

Biological systems perceive the world by simultaneously processing high-dimensional inputs from modalities as diverse as vision, audition, touch, proprioception, etc. The perception models used in deep learning on the other hand are designed for individual modalities, often relying on domain-specific assumptions such as the local grid structures exploited by virtually all existing vision models. These priors introduce helpful inductive biases, but also lock models to individual modalities. In this paper we introduce the Perceiver - a model that builds upon Transformers and hence makes few architectural assumptions about the relationship between its inputs, but that also scales to hundreds of thousands of inputs, like ConvNets. The model leverages an asymmetric attention mechanism to iteratively distill inputs into a tight latent bottleneck, allowing it to scale to handle very large inputs. We show that this architecture is competitive with or outperforms strong, specialized models on classification tasks across various modalities: images, point clouds, audio, video, and video+audio. The Perceiver obtains performance comparable to ResNet-50 and ViT on ImageNet without 2D convolutions by directly attending to 50,000 pixels. It is also competitive in all modalities in AudioSet.

In this paper, we propose Image Downscaling Assessment by Rate-Distortion (IDA-RD), a novel measure to quantitatively evaluate image downscaling algorithms. In contrast to image-based methods that measure the quality of downscaled images, ours is process-based that draws ideas from rate-distortion theory to measure the distortion incurred during downscaling. Our main idea is that downscaling and super-resolution (SR) can be viewed as the encoding and decoding processes in the rate-distortion model, respectively, and that a downscaling algorithm that preserves more details in the resulting low-resolution (LR) images should lead to less distorted high-resolution (HR) images in SR. In other words, the distortion should increase as the downscaling algorithm deteriorates. However, it is non-trivial to measure this distortion as it requires the SR algorithm to be blind and stochastic. Our key insight is that such requirements can be met by recent SR algorithms based on deep generative models that can find all matching HR images for a given LR image on their learned image manifolds. Extensive experimental results show the effectiveness of our IDA-RD measure.

Understanding the quality of a performance evaluation metric is crucial for ensuring that model outputs align with human preferences. However, it remains unclear how well each metric captures the diverse aspects of these preferences, as metrics often excel in one particular area but not across all dimensions. To address this, it is essential to systematically calibrate metrics to specific aspects of human preference, catering to the unique characteristics of each aspect. We introduce MetaMetrics, a calibrated meta-metric designed to evaluate generation tasks across different modalities in a supervised manner. MetaMetrics optimizes the combination of existing metrics to enhance their alignment with human preferences. Our metric demonstrates flexibility and effectiveness in both language and vision downstream tasks, showing significant benefits across various multilingual and multi-domain scenarios. MetaMetrics aligns closely with human preferences and is highly extendable and easily integrable into any application. This makes MetaMetrics a powerful tool for improving the evaluation of generation tasks, ensuring that metrics are more representative of human judgment across diverse contexts.

As the focus in LLM-based coding shifts from static single-step code generation to multi-step agentic interaction with tools and environments, understanding which tasks will challenge agents and why becomes increasingly difficult. This is compounded by current practice: agent performance is typically measured by aggregate pass rates on benchmarks, but single-number metrics obscure the diversity of tasks within a benchmark. We present a framework for predicting success or failure on individual tasks tailored to the agentic coding regime. Our approach augments Item Response Theory (IRT) with rich features extracted from tasks, including issue statements, repository contexts, solutions, and test cases, and introduces a novel decomposition of agent ability into LLM and scaffold ability components. This parameterization enables us to aggregate evaluation data across heterogeneous leaderboards and accurately predict task-level performance for unseen benchmarks, as well as unseen LLM-scaffold combinations. Our methods have practical utility for benchmark designers, who can better calibrate the difficulty of their new tasks without running computationally expensive agent evaluations.

Human subjective evaluation is the gold standard to evaluate speech quality optimized for human perception. Perceptual objective metrics serve as a proxy for subjective scores. The conventional and widely used metrics require a reference clean speech signal, which is unavailable in real recordings. The no-reference approaches correlate poorly with human ratings and are not widely adopted in the research community. One of the biggest use cases of these perceptual objective metrics is to evaluate noise suppression algorithms. This paper introduces a multi-stage self-teaching based perceptual objective metric that is designed to evaluate noise suppressors. The proposed method generalizes well in challenging test conditions with a high correlation to human ratings.

RL for code correctness is now established: have the model generate a program, run it against hidden test cases, and reward solutions that pass. Extending this to code optimization seems straightforward: just add execution time to the reward. But in practice, once timing drives the reward, small problems in measurement noise, reward sparsity, or GRPO instability overwhelm the signal and make RL fail: generated solutions are barely faster, and more of them can fail. We make execution time learnable through three stages: (1) how code is tested, by building DMC-Optim with large optimization tests and a calibrated sandbox; (2) how speed is turned into reward, by composing correctness and speed in the RL environment and using an offline simulator to predict the most promising configurations; and (3) how the model learns from that reward, by adapting GRPO and evaluation to the sparser, noisier timed-execution setting. On DMC-Optim, the strongest optimization-aware configurations improve strict top-50% pass@1 from 18.0% to 31.3% on Qwen 2.5 7B and from 30.7% to 50.4% on CWM 32B. These gains further increase at stricter percentiles such as top-30%, with 125% relative improvement for CWM 32B, while preserving pure-correctness scores. When the timing sandbox is degraded, robust optimization RL reaches 100% to 200% improvement over standard RLVR, depending on the evaluation criterion. On LCB, CWM 32B wins up to 83% of median-sample speed comparisons against standard RLVR. Relative to the fastest correct human submissions per problem, it reaches about half the human rate of complexity-class improvements (14% vs. 28%).

Rate control algorithms are at the heart of video conferencing platforms, determining target bitrates that match dynamic network characteristics for high quality. Recent data-driven strategies have shown promise for this challenging task, but the performance degradation they introduce during training has been a nonstarter for many production services, precluding adoption. This paper aims to bolster the practicality of data-driven rate control by presenting an alternative avenue for experiential learning: leveraging purely existing telemetry logs produced by the incumbent algorithm in production. We observe that these logs contain effective decisions, although often at the wrong times or in the wrong order. To realize this approach despite the inherent uncertainty that log-based learning brings (i.e., lack of feedback for new decisions), our system, Mowgli, combines a variety of robust learning techniques (i.e., conservatively reasoning about alternate behavior to minimize risk and using a richer model formulation to account for environmental noise). Across diverse networks (emulated and real-world), Mowgli outperforms the widely deployed GCC algorithm, increasing average video bitrates by 15-39% while reducing freeze rates by 60-100%.

Reinforcement Learning from Human Feedback (RLHF) involves training policy models (PMs) and reward models (RMs) to align language models with human preferences. Instead of focusing solely on PMs and RMs independently, we propose to examine their interactions during fine-tuning, introducing the concept of seamlessness. Our study starts with observing the saturation phenomenon, where continual improvements in RM and PM do not translate into RLHF progress. Our analysis shows that RMs fail to assign proper scores to PM responses, resulting in a 35% mismatch rate with human preferences, highlighting a significant discrepancy between PM and RM. To measure seamlessness between PM and RM without human effort, we propose an automatic metric, SEAM. SEAM quantifies the discrepancies between PM and RM judgments induced by data samples. We validate the effectiveness of SEAM in data selection and model augmentation. Our experiments demonstrate that (1) using SEAM-filtered data for RL training improves RLHF performance by 4.5%, and (2) SEAM-guided model augmentation results in a 4% performance improvement over standard augmentation methods.

Recent large audio language models (LALMs) have extended their capabilities beyond lexical speech understanding to tasks requiring paralinguistic and non-linguistic audio understanding, such as emotion recognition. However, these models still struggle to reliably utilize acoustic information, particularly when it co-occurs with spoken linguistic content. Existing benchmarks do not clearly distinguish acoustic perception from higher-level reasoning, making it difficult to identify the source of these failures. To address this issue, we introduce VocalGrad, a benchmark designed to evaluate LALMs' ability to perceive low-level acoustic variations. VocalGrad formulates this problem as a simple binary classification of temporal change direction, enabling evaluation largely independent of higher-level reasoning or semantic understanding. Our benchmark results reveal a substantial gap in acoustic perception between human annotators and state-of-the-art LALMs: while human annotators achieve high accuracy, LALMs perform at near-chance levels. We further conduct linear probing to identify structural bottlenecks in these models. Our analysis reveals two distinct failure modes: for some categories, task-relevant information is preserved in internal representations but not utilized by the language model head; for others, the information becomes progressively less linearly decodable across model layers. Overall, our results reveal substantial limitations of LALMs in modeling acoustic variation under controlled settings and expose representation differences not observable from text outputs.

The scientific scale-up of large language models (LLMs) necessitates a comprehensive understanding of their scaling properties. However, the existing literature on the scaling properties only yields an incomplete answer: optimization loss decreases predictably as the model size increases, in line with established scaling law; yet no scaling law for task has been established and the task performances are far from predictable during scaling. Task performances typically show minor gains on small models until they improve dramatically once models exceed a size threshold, exemplifying the ``emergent abilities''. In this study, we discover that small models, although they exhibit minor performance, demonstrate critical and consistent task performance improvements that are not captured by conventional evaluation strategies due to insufficient measurement resolution. To measure such improvements, we introduce PassUntil, an evaluation strategy through massive sampling in the decoding phase. We conduct quantitative investigations into the scaling law of task performance. Firstly, a strict task scaling law is identified, enhancing the predictability of task performances. Remarkably, we are able to predict the performance of the 2.4B model on code generation with merely 0.05\% deviation before training starts. Secondly, underpinned by PassUntil, we observe concrete evidence of emergent abilities and ascertain that they are not in conflict with the continuity of performance improvement. Their semblance to break-through is that their scaling curve cannot be fitted by standard scaling law function. We then introduce a mathematical definition for the emergent abilities. Through the definition, we refute a prevalent ``multi-step reasoning hypothesis'' regarding the genesis of emergent abilities and propose a new hypothesis with a satisfying fit to the observed scaling curve.

Omni-proactive streaming video understanding, i.e., autonomously deciding when to speak and what to say from continuous audio-visual streams, is an emerging capability of omni-modal large language models. Existing benchmarks fall short in three key aspects: they rely primarily on visual signals, adopt polling or fixed-timestamp protocols instead of true proactive evaluation, and cover only a limited range of tasks, preventing reliable assessment and differentiation of omni-proactive streaming models. We present OmniPro, the first benchmark to jointly evaluate omni-modal perception, proactive responding, and diverse video understanding tasks. It comprises 2,700 human-verified samples spanning 9 sub-tasks and 3 cognitive levels, covering 6 basic video understanding capabilities. Notably, 84% of samples require audio signals (speech or non-speech), and each sample is annotated with modality-isolation labels to enable fine-grained multimodal analysis. We further introduce a dual-mode evaluation protocol: Probe mode assesses content understanding by querying the model before and after each ground-truth trigger, while Online mode evaluates full proactive ability by requiring models to autonomously decide when to respond in streaming input. Evaluating 11 representative models reveals three key findings: (1) audio provides consistent gains but with highly variable utilization across models, (2) performance degrades significantly over time, indicating limited long-horizon robustness, and (3) non-speech audio perception remains the weakest dimension.

Despite rapid progress in video-capable MLLMs, we find that their apparent audio understanding in videos is often vision-driven: models rely on visual cues to infer or hallucinate acoustic information, rather than verifying the audio stream. This issue appears across both state-of-the-art open-source omni models and leading closed-source models from providers such as Google and OpenAI. We characterize this failure mode as an audio-visual Clever Hans effect, in which models appear (falsely) audio-grounded, but actually exploit visual-acoustic correlations without verifying whether the audio and visual streams are truly aligned. To systematically study this behavior, we introduce Thud, an intervention-driven probing framework based on three counterfactual audio edits: Shift, which tests temporal synchronization; Mute, which tests sound existence; and Swap, which tests audio-visual consistency. Beyond diagnosis, we further study a two-stage alignment recipe: intervention-derived preference pairs teach audio verification, while event-level general video preferences regularize the model against over-specialization. Our best 10K-sample recipe improves average performance across the three intervention dimensions by 28 percentage points, while slightly improving performance on general video and audio-visual QA benchmarks.

Instruction-tuning data are judged by quality metrics, and tuned models are judged by benchmarks, but both judgments pass through an output interface: the surface format in which an answer is written. Using gradient signatures across 12 tasks, four semantically equivalent interfaces, three model families, and controlled corruptions, we show that this interface confounds both measurements. Spectral statistics such as effective rank are provably invariant to interface rotation and empirically blind to semantic corruption, while the direction of the update carries the quality signal. The interface-varying residual is not noise: it identifies each unit's own target task perfectly across all three families. Capability itself is stored relative to the training interface: a skill that raises accuracy by more than 40 points under the training format can be nearly invisible under every other, and correcting a single generation budget flips the measured effect of fine-tuning on GSM8K from a gain into a large loss. Pre-registered interventions delimit where this geometry stops short of control. Data quality and model capability are interface-conditioned quantities, and current practice often reports the interface instead of the content.

The rise of machine-learning systems that process sensory input has brought with it a rise in comparisons between human and machine perception. But such comparisons face a challenge: Whereas machine perception of some stimulus can often be probed through direct and explicit measures, much of human perceptual knowledge is latent, incomplete, or unavailable for explicit report. Here, we explore how this asymmetry can cause such comparisons to misestimate the overlap in human and machine perception. As a case study, we consider human perception of adversarial speech -- synthetic audio commands that are recognized as valid messages by automated speech-recognition systems but that human listeners reportedly hear as meaningless noise. In five experiments, we adapt task designs from the human psychophysics literature to show that even when subjects cannot freely transcribe such speech commands (the previous benchmark for human understanding), they often can demonstrate other forms of understanding, including discriminating adversarial speech from closely matched non-speech (Experiments 1--2), finishing common phrases begun in adversarial speech (Experiments 3--4), and solving simple math problems posed in adversarial speech (Experiment 5) -- even for stimuli previously described as unintelligible to human listeners. We recommend the adoption of such "sensitive tests" when comparing human and machine perception, and we discuss the broader consequences of such approaches for assessing the overlap between systems.

Coding agents are now commonly evaluated on the SWE-bench family of benchmarks, whose tasks are built from curated GitHub issues: long, structured, and information-rich. Real user requests, however, are typically far shorter and less structured. To characterize this gap, we define a six-category information taxonomy and four dimensions of linguistic style, and apply them to real user prompts from SWE-chat and problem statements from SWE-bench Verified and Pro. We find that requests carrying only a problem statement, alone or with limited additional context, account for 88% of real prompts but just 7% of benchmark problems. Furthermore, 87% of real prompts are casually written whereas 94% of benchmark problems are formal. Guided by these observations, we introduce RealSWE, 381 multi-variant task families derived from SWE-bench Verified and Pro. Variants within each family share the same underlying task and gold patch while differing only in information composition and linguistic style. Evaluating seven contemporary LLMs with RealSWE, we find that i) realistic inputs reduce resolution rates by 6.4 pp on average and can change model rankings. Controlled analysis further shows that ii) including Desired Behavior and Motivation significantly affects performance, whereas Environment Information and Reproduction Steps merely add tokens without measurable benefit; iii) linguistic style has only small, model-dependent effects. These findings provide actionable guidance for users and agents: explicitly stating the desired behavior and motivation, which most real prompts omit, substantially improves the LLM's software engineering performance.

The ability of large vision-language models (LVLMs) to critique and correct their reasoning is an essential building block towards their self-improvement. However, a systematic analysis of such capabilities in LVLMs is still lacking. We propose VISCO, the first benchmark to extensively analyze the fine-grained critique and correction capabilities of LVLMs. Compared to existing work that uses a single scalar value to critique the entire reasoning [4], VISCO features dense and fine-grained critique, requiring LVLMs to evaluate the correctness of each step in the chain-of-thought and provide natural language explanations to support their judgments. Extensive evaluation of 24 LVLMs demonstrates that human-written critiques significantly enhance the performance after correction, showcasing the potential of the self-improvement strategy. However, the model-generated critiques are less helpful and sometimes detrimental to the performance, suggesting that critique is the crucial bottleneck. We identified three common patterns in critique failures: failure to critique visual perception, reluctance to "say no", and exaggerated assumption of error propagation. To address these issues, we propose an effective LookBack strategy that revisits the image to verify each piece of information in the initial reasoning. LookBack significantly improves critique and correction performance by up to 13.5%.

Reinforcement Learning with Verifiable Rewards (RLVR) has proven to be a highly effective strategy for endowing Large Language Models (LLMs) with robust multi-step reasoning abilities. However, its design and optimizations remain tailored to purely textual domains, resulting in suboptimal performance when applied to multimodal reasoning tasks. In particular, we observe that a major source of error in current multimodal reasoning lies in the perception of visual inputs. To address this bottleneck, we propose Perception-Aware Policy Optimization (PAPO), a simple yet effective extension of GRPO that encourages the model to learn to perceive while learning to reason, entirely from internal supervision signals. Notably, PAPO does not rely on additional data curation, external reward models, or proprietary models. Specifically, we introduce the Implicit Perception Loss in the form of a KL divergence term to the GRPO objective, which, despite its simplicity, yields significant overall improvements (4.4%) on diverse multimodal benchmarks. The improvements are more pronounced, approaching 8.0%, on tasks with high vision dependency. We also observe a substantial reduction (30.5%) in perception errors, indicating improved perceptual capabilities with PAPO. We conduct comprehensive analysis of PAPO and identify a unique loss hacking issue, which we rigorously analyze and mitigate through a Double Entropy Loss. Overall, our work introduces a deeper integration of perception-aware supervision into RLVR learning objectives and lays the groundwork for a new RL framework that encourages visually grounded reasoning. Project page: https://mikewangwzhl.github.io/PAPO.

Large language models excel at reasoning but lack key aspects of introspection, including anticipating their own success and the computation required to achieve it. Humans use real-time introspection to decide how much effort to invest, when to make multiple attempts, when to stop, and when to signal success or failure. Without this, LLMs struggle to make intelligent meta-cognition decisions. Test-time scaling methods like Best-of-N drive up cost and latency by using a fixed budget of samples regardless of the marginal benefit of each one at any point in generation, and the absence of confidence signals can mislead people, prevent appropriate escalation to better tools, and undermine trustworthiness. Learned verifiers or reward models can provide confidence estimates, but do not enable adaptive inference and add substantial cost by requiring extra models or forward passes. We present ZIP-RC, which equips models with zero-overhead introspective predictions of reward and cost. At every token, ZIP-RC reuses reserved or unused logits in the same forward pass as next-token prediction to output a joint distribution over final reward and remaining length -- no extra models, architecture change, or inference overhead. This full joint distribution is used to compute a sampling utility which is the linear combination of the expected maximum reward, total compute, and latency of set of samples if generated to completion. During inference, we maximize this utility with meta-actions that determine which prefix of tokens to continue or initiate sampling from. On mixed-difficulty mathematical benchmarks, ZIP-RC improves accuracy by up to 12% over majority voting at equal or lower average cost, and traces smooth Pareto frontiers between quality, compute, and latency. By providing real-time reward-cost introspection, ZIP-RC enables adaptive, efficient reasoning.

In arena-style evaluation of large language models (LLMs), two LLMs respond to a user query, and the user chooses the winning response or deems the "battle" a draw, resulting in an adjustment to the ratings of both models. The prevailing approach for modeling these rating dynamics is to view battles as two-player game matches, as in chess, and apply the Elo rating system and its derivatives. In this paper, we critically examine this paradigm. Specifically, we question whether a draw genuinely means that the two models are equal and hence whether their ratings should be equalized. Instead, we conjecture that draws are more indicative of query difficulty: if the query is too easy, then both models are more likely to succeed equally. On three real-world arena datasets, we show that ignoring rating updates for draws yields a 1-3% relative increase in battle outcome prediction accuracy (which includes draws) for all four rating systems studied. Further analyses suggest that draws occur more for queries rated as very easy and those as highly objective, with risk ratios of 1.37 and 1.35, respectively. We recommend future rating systems to reconsider existing draw semantics and to account for query properties in rating updates.

Real-time interaction requires agents to perceive evolving environments and act under temporal constraints, yet existing evaluations of real-time streaming models focus on conversational quality and streaming perception, leaving effective action under strict temporal constraints untested. We introduce Realtime-Bench, a suite of carefully curated dynamic tasks where delayed actions directly affect success. Under Wall Clock, where environments advance according to actual inference latency, the best success rate reaches only 44% and the best Pass^3 is only 36%, demonstrating that real-time interaction remains challenging. To separate serving latency from models' real-time streaming capability, we compare Wall Clock with Paused Clock, which removes inference delays as an idealized reference, and introduce Token Clock, which advances simulated time according to generated token counts, testing temporal awareness independently of physical serving speed. We find that temporal costs limit success in evolving environments and that higher reasoning effort has mixed effects under time pressure, reflecting a trade-off between decision quality and timely execution. These findings suggest directions for improving real-time multimodal agents, including adaptive reasoning with temporal awareness and continuous perception and action.

Inspired by the success of DeepSeek-R1, we explore the potential of rule-based reinforcement learning (RL) in MLLM post-training for perception policy learning. While promising, our initial experiments reveal that incorporating a thinking process through RL does not consistently lead to performance gains across all visual perception tasks. This leads us to delve into the essential role of RL in the context of visual perception. In this work, we return to the fundamentals and explore the effects of RL on different perception tasks. We observe that the perceptual complexity is a major factor in determining the effectiveness of RL. We also observe that reward design plays a crucial role in further approching the upper limit of model perception. To leverage these findings, we propose Perception-R1, a scalable RL framework using GRPO during MLLM post-training. With a standard Qwen2.5-VL-3B-Instruct, Perception-R1 achieves +4.2% on RefCOCO+, +17.9% on PixMo-Count, +4.2% on PageOCR, and notably, 31.9% AP on COCO2017 val for the first time, establishing a strong baseline for perception policy learning.

Multimodal large language models (MLLMs) often struggle to ground reasoning in perceptual evidence. We present a systematic study of perception strategies-explicit, implicit, visual, and textual-across four multimodal benchmarks and two MLLMs. Our findings show that explicit perception, especially when paired with textual cues, consistently yields the best improvements, particularly for smaller models. Based on this insight, we propose VTPerception-R1, a unified two-stage framework that decouples perception from reasoning. Stage 1 introduces perception-augmented fine-tuning, and Stage 2 applies perception-aware reinforcement learning with novel visual, textual, and consistency rewards. Experiments demonstrate that VTPerception-R1 significantly improves reasoning accuracy and robustness across diverse tasks, offering a scalable and auditable solution for perception-grounded multimodal reasoning. Our code is available at: https://github.com/yizhuoDi/VTPerceprion-R1.

This paper studies contextual biasing with Large Language Models (LLMs), where during second-pass rescoring additional contextual information is provided to a LLM to boost Automatic Speech Recognition (ASR) performance. We propose to leverage prompts for a LLM without fine tuning during rescoring which incorporate a biasing list and few-shot examples to serve as additional information when calculating the score for the hypothesis. In addition to few-shot prompt learning, we propose multi-task training of the LLM to predict both the entity class and the next token. To improve the efficiency for contextual biasing and to avoid exceeding LLMs' maximum sequence lengths, we propose dynamic prompting, where we select the most likely class using the class tag prediction, and only use entities in this class as contexts for next token prediction. Word Error Rate (WER) evaluation is performed on i) an internal calling, messaging, and dictation dataset, and ii) the SLUE-Voxpopuli dataset. Results indicate that biasing lists and few-shot examples can achieve 17.8% and 9.6% relative improvement compared to first pass ASR, and that multi-task training and dynamic prompting can achieve 20.0% and 11.3% relative WER improvement, respectively.

Evaluations of image compression performance which include human preferences have generally found that naive distortion functions such as MSE are insufficiently aligned to human perception. In order to align compression models to human perception, prior work has employed differentiable perceptual losses consisting of neural networks calibrated on large-scale datasets of human psycho-visual judgments. We show that, surprisingly, state-of-the-art vision-language models (VLMs) can replicate binary human two-alternative forced choice (2AFC) judgments zero-shot when asked to reason about the differences between pairs of images. Motivated to exploit the powerful zero-shot visual reasoning capabilities of VLMs, we propose Vision-Language Models for Image Compression (VLIC), a diffusion-based image compression system designed to be post-trained with binary VLM judgments. VLIC leverages existing techniques for diffusion model post-training with preferences, rather than distilling the VLM judgments into a separate perceptual loss network. We show that calibrating this system on VLM judgments produces competitive or state-of-the-art performance on human-aligned visual compression depending on the dataset, according to perceptual metrics and large-scale user studies. We additionally conduct an extensive analysis of the VLM-based reward design and training procedure and share important insights. More visuals are available at https://kylesargent.github.io/vlic

AI agents in partially observable environments need to coordinate active sensing with working memory to maintain an evolving perceptual state. However, existing benchmarks struggle to isolate this perceptual-state construction and interpretation capability because they introduce physical and control complexities. We address this with MNIST-PRO, a benchmark that isolates agentic perception by converting MNIST digit recognition into a sequential, glimpse-based search task with lookback constraints. We evaluate ten multimodal models across four memory representations, including raw visual history, textual states, structured metric grid maps, and a consolidated visual canvas. While models excel under full observability, partial observability exposes a clear performance gap. We identify three distinct bottlenecks. First, perceptual-state construction and interpretation present a challenge, as agents struggle to integrate fragmented glimpses. Second, agents often stop exploring before they see the full sequence. Third, models often fail to revise early, incorrect beliefs even when faced with subsequent contradictory evidence. These results show that simply acquiring visual evidence is not enough. Agents must also be able to build and update a reliable perceptual state.

Performance optimization is a critical yet challenging aspect of software development, often requiring a deep understanding of system behavior, algorithmic tradeoffs, and careful code modifications. Although recent advances in AI coding agents have accelerated code generation and bug fixing, little is known about how these agents perform on real-world performance optimization tasks. We present the first empirical study comparing agent- and human-authored performance optimization commits, analyzing 324 agent-generated and 83 human-authored PRs from the AIDev dataset across adoption, maintainability, optimization patterns, and validation practices. We find that AI-authored performance PRs are less likely to include explicit performance validation than human-authored PRs (45.7\% vs. 63.6\%, p=0.007). In addition, AI-authored PRs largely use the same optimization patterns as humans. We further discuss limitations and opportunities for advancing agentic code optimization.

Within the ambit of VoIP (Voice over Internet Protocol) telecommunications, the complexities introduced by acoustic transformations merit rigorous analysis. This research, rooted in the exploration of proprietary sender-side denoising effects, meticulously evaluates platforms such as Google Meets and Zoom. The study draws upon the Deep Noise Suppression (DNS) 2020 dataset, ensuring a structured examination tailored to various denoising settings and receiver interfaces. A methodological novelty is introduced via Blinder-Oaxaca decomposition, traditionally an econometric tool, repurposed herein to analyze acoustic-phonetic perturbations within VoIP systems. To further ground the implications of these transformations, psychoacoustic metrics, specifically PESQ and STOI, were used to explain of perceptual quality and intelligibility. Cumulatively, the insights garnered underscore the intricate landscape of VoIP-influenced acoustic dynamics. In addition to the primary findings, a multitude of metrics are reported, extending the research purview. Moreover, out-of-domain benchmarking for both time and time-frequency domain speech enhancement models is included, thereby enhancing the depth and applicability of this inquiry.

Large language models (LLMs) are widely used as zero-shot and few-shot classifiers, where task behaviour is largely controlled through prompting. A growing number of works have observed that LLMs are sensitive to prompt variations, with small changes leading to large changes in performance. However, in many cases, the investigation of sensitivity is performed using underspecified prompts that provide minimal task instructions and weakly constrain the model's output space. In this work, we argue that a significant portion of the observed prompt sensitivity can be attributed to prompt underspecification. We systematically study and compare the sensitivity of underspecified prompts and prompts that provide specific instructions. Utilising performance analysis, logit analysis, and linear probing, we find that underspecified prompts exhibit higher performance variance and lower logit values for relevant tokens, while instruction-prompts suffer less from such problems. However, linear probing analysis suggests that the effects of prompt underspecification have only a marginal impact on the internal LLM representations, instead emerging in the final layers. Overall, our findings highlight the need for more rigour when investigating and mitigating prompt sensitivity.

Current benchmarks for assessing vision-language models (VLMs) often focus on their perception or problem-solving capabilities and neglect other critical aspects such as fairness, multilinguality, or toxicity. Furthermore, they differ in their evaluation procedures and the scope of the evaluation, making it difficult to compare models. To address these issues, we extend the HELM framework to VLMs to present the Holistic Evaluation of Vision Language Models (VHELM). VHELM aggregates various datasets to cover one or more of the 9 aspects: visual perception, knowledge, reasoning, bias, fairness, multilinguality, robustness, toxicity, and safety. In doing so, we produce a comprehensive, multi-dimensional view of the capabilities of the VLMs across these important factors. In addition, we standardize the standard inference parameters, methods of prompting, and evaluation metrics to enable fair comparisons across models. Our framework is designed to be lightweight and automatic so that evaluation runs are cheap and fast. Our initial run evaluates 22 VLMs on 21 existing datasets to provide a holistic snapshot of the models. We uncover new key findings, such as the fact that efficiency-focused models (e.g., Claude 3 Haiku or Gemini 1.5 Flash) perform significantly worse than their full models (e.g., Claude 3 Opus or Gemini 1.5 Pro) on the bias benchmark but not when evaluated on the other aspects. For transparency, we release the raw model generations and complete results on our website (https://crfm.stanford.edu/helm/vhelm/v2.0.1). VHELM is intended to be a living benchmark, and we hope to continue adding new datasets and models over time.

Human feedback has become the de facto standard for evaluating the performance of Large Language Models, and is increasingly being used as a training objective. However, it is not clear which properties of a generated output this single `preference' score captures. We hypothesise that preference scores are subjective and open to undesirable biases. We critically analyse the use of human feedback for both training and evaluation, to verify whether it fully captures a range of crucial error criteria. We find that while preference scores have fairly good coverage, they under-represent important aspects like factuality. We further hypothesise that both preference scores and error annotation may be affected by confounders, and leverage instruction-tuned models to generate outputs that vary along two possible confounding dimensions: assertiveness and complexity. We find that the assertiveness of an output skews the perceived rate of factuality errors, indicating that human annotations are not a fully reliable evaluation metric or training objective. Finally, we offer preliminary evidence that using human feedback as a training objective disproportionately increases the assertiveness of model outputs. We encourage future work to carefully consider whether preference scores are well aligned with the desired objective.

As large language models (LLMs) are adopted as a fundamental component of language technologies, it is crucial to accurately characterize their performance. Because choices in prompt design can strongly influence model behavior, this design process is critical in effectively using any modern pre-trained generative language model. In this work, we focus on LLM sensitivity to a quintessential class of meaning-preserving design choices: prompt formatting. We find that several widely used open-source LLMs are extremely sensitive to subtle changes in prompt formatting in few-shot settings, with performance differences of up to 76 accuracy points when evaluated using LLaMA-2-13B. Sensitivity remains even when increasing model size, the number of few-shot examples, or performing instruction tuning. Our analysis suggests that work evaluating LLMs with prompting-based methods would benefit from reporting a range of performance across plausible prompt formats, instead of the currently-standard practice of reporting performance on a single format. We also show that format performance only weakly correlates between models, which puts into question the methodological validity of comparing models with an arbitrarily chosen, fixed prompt format. To facilitate systematic analysis we propose FormatSpread, an algorithm that rapidly evaluates a sampled set of plausible prompt formats for a given task, and reports the interval of expected performance without accessing model weights. Furthermore, we present a suite of analyses that characterize the nature of this sensitivity, including exploring the influence of particular atomic perturbations and the internal representation of particular formats.

As multimodal large language models (MLLMs) advance rapidly, rigorous evaluation has become essential, providing further guidance for their development. In this work, we focus on a unified and robust evaluation of vision perception abilities, the foundational skill of MLLMs. We find that existing perception benchmarks, each focusing on different question types, domains, and evaluation metrics, introduce significant evaluation variance, complicating comprehensive assessments of perception abilities when relying on any single benchmark. To address this, we introduce AbilityLens, a unified benchmark designed to evaluate MLLMs across six key perception abilities, focusing on both accuracy and stability, with each ability encompassing diverse question types, domains, and metrics. With the assistance of AbilityLens, we: (1) identify the strengths and weaknesses of current models, highlighting stability patterns and revealing a notable performance gap between open-source and closed-source models; (2) introduce an online evaluation mode, which uncovers interesting ability conflict and early convergence phenomena during MLLM training; and (3) design a simple ability-specific model merging method that combines the best ability checkpoint from early training stages, effectively mitigating performance decline due to ability conflict. The benchmark and online leaderboard will be released soon.

Code models are increasingly trained with execution feedback, but most training signals still stop at correctness. This leaves an important gap for systems code: two programs can pass the same tests while differing greatly in runtime. We study how to train code agents to prefer faster correct implementations, rather than treating efficiency only as an evaluation metric. The key difficulty is that runtime is a fragile reward. It is meaningful only after a program is correct, varies across tasks, and gives little guidance when most sampled programs fail to compile or run. We propose RLPF, reinforcement learning from performance feedback, which turns execution outcomes into a staged reward. Failed programs are ordered by execution progress, while correct programs are ranked by their relative improvement from the baseline toward the expert reference. This gives useful feedback before correctness and performance-sensitive feedback after correctness. Fine-tuning Qwen3-32B with RLPF on PerfCodeBench raises correct-and-runnable solutions from 11.1% to 54.6% and improves relative efficiency from 8.1% to 38.6%. The trained model becomes competitive with stronger open-weight systems, and its optimization behavior transfers modestly to EffiBench-X. Additional studies show that model-generated references provide useful but weaker supervision, and that the full composite reward is more reliable than correctness-only or runtime-only baselines. These results suggest that code agents can be trained not only to pass tests, but also to optimize the programs they write.

When audio, video, and text disagree, accuracy alone does not show whether to acquire another source or abstain. We study these choices in a controlled three-score benchmark: a policy observes two signed scores, may request the third at a cost, and can abstain. The primary reward is mechanism-specific: abstention is correct only for one designated ambiguity mechanism and is penalized under mixed corruption. Matched controls show that a threshold policy matches always-request decisions with fewer requests; its advantage over always-answer fusion depends on the reward assigned to that ambiguity. On a partially held-out synthetic split, the threshold policy reaches 0.789 +/- 0.006 targeted decision accuracy and 0.481 +/- 0.014 utility across 83 seeds. A three-score majority reference reaches 0.626 +/- 0.008 and 0.252 +/- 0.016, but uses more information. In a matched-budget test, a train-only value selector improves utility over no-query and matched-random policies at 10% and 25% budgets, while pair uncertainty has higher utility at every budget. At 50% and 63.7% budgets, the selector lowers utility despite slightly higher non-ambiguous accuracy. If all abstentions are scored incorrect, majority outranks the threshold policy in utility. At a central temporal setting, full-trace controls match the neural models while position perturbations separate them. On held-out-actor emotion clips, eight-frame fusion has opposite-signed accuracy differences for two encoder pairs, with both actor intervals containing zero; matched-request routing gains are small and uncertain. These results separate full-modality accuracy from pre-request selection value and show that selection value depends on budget and the observed-pair ranking.

Multimodal Large Language Models (MLLMs) are increasingly deployed in human-facing roles where personality perception is critical, yet existing benchmarks evaluate this capability solely on numerical Big Five score prediction, leaving open whether models truly perceive personality through behavioral understanding or merely prejudge through superficial pattern matching. We address this gap with three contributions. (i) A new task: we formalize Grounded Personality Reasoning (GPR), which requires MLLMs to anchor each Big Five rating in observable evidence through a chain of rating, reasoning, and grounding. (ii) A new dataset: we release MM-OCEAN (1,104 videos, 5,320 MCQs), produced by a multi-agent pipeline with human verification, with timestamped behavioral observations, evidence-grounded trait analyses, and seven categories of cue-grounding MCQs. (iii) Benchmark and analysis: we design a three-tier evaluation (rating, reasoning, grounding) plus four sample-level failure-mode metrics: Prejudice Rate (PR), Confabulation Rate (CR), Integration-failure Rate (IR), and Holistic-grounding Rate (HR), and benchmark 27 MLLMs (13 closed, 14 open). The analysis uncovers a striking Prejudice Gap: across the field, 51% of correct ratings are not grounded in retrieved cues, and the Holistic-Grounding Rate spans only 0-33.5%. These findings expose a disconnect between getting the right score and reasoning for the right reason, charting a roadmap for grounded social cognition in MLLMs.

Large language models (LLMs) are increasingly adopted in educational technologies for a variety of tasks, from generating instructional materials and assisting with assessment design to tutoring. While prior work has investigated how models can be adapted or optimized for specific tasks, far less is known about how well LLMs perform at interpreting authentic educational scenarios without significant customization. As LLM-based systems become widely adopted by learners and educators in everyday academic contexts, understanding their out-of-the-box capabilities is increasingly important for setting expectations and benchmarking. We compared six LLMs to estimate their baseline performance on a simple but important task: classifying instructional moves in authentic classroom transcripts. We evaluated typical prompting methods: zero-shot, one-shot, and few-shot prompting. We found that while zero-shot performance was moderate, providing comprehensive examples (few-shot prompting) significantly improved performance for state-of-the-art models, with the strongest configuration reaching Cohen's Kappa = 0.58 against expert-coded annotations. At the same time, improvements were neither uniform nor complete: performance varied considerably by instructional move, and higher recall frequently came at the cost of increased false positives. Overall, these findings indicate that foundation models demonstrate meaningful yet limited capacity to interpret instructional discourse, with prompt design helping to surface capability but not eliminating fundamental reliability constraints.

Large Audio Language Models (LALMs) achieve strong performance on a variety of audio understanding tasks but continue to struggle with temporal reasoning, a fundamental capability central to human auditory perception. Understanding the causes of these failures remains challenging as existing benchmarks report performance gaps without probing underlying mechanisms. To address this, we introduce a benchmark with 1,657 questions across three foundational tasks designed specifically for mechanistic analysis. Examining model outputs across varying input settings (behavioral analysis) reveals that models often under-utilize audio when textual cues are available. We also provide the first causal mechanistic analysis of temporal reasoning failures in LALMs. Comparing attention upweighting against scaling, we find that redistributing attention across audio tokens is more effective than increasing audio attention. Targeting task-relevant tokens yields further gains. These findings suggest that modality imbalance alone cannot explain failures. Attention scaling at bottleneck layers improves accuracy from 55.9% to 59.1% without fine-tuning, demonstrating a promising direction for future work.

In this paper, we evaluate different abilities of GPT-4V including visual understanding, language understanding, visual puzzle solving, and understanding of other modalities such as depth, thermal, video, and audio. To estimate GPT-4V's performance, we manually construct 656 test instances and carefully evaluate the results of GPT-4V. The highlights of our findings are as follows: (1) GPT-4V exhibits impressive performance on English visual-centric benchmarks but fails to recognize simple Chinese texts in the images; (2) GPT-4V shows inconsistent refusal behavior when answering questions related to sensitive traits such as gender, race, and age; (3) GPT-4V obtains worse results than GPT-4 (API) on language understanding tasks including general language understanding benchmarks and visual commonsense knowledge evaluation benchmarks; (4) Few-shot prompting can improve GPT-4V's performance on both visual understanding and language understanding; (5) GPT-4V struggles to find the nuances between two similar images and solve the easy math picture puzzles; (6) GPT-4V shows non-trivial performance on the tasks of similar modalities to image, such as video and thermal. Our experimental results reveal the ability and limitations of GPT-4V and we hope our paper can provide some insights into the application and research of GPT-4V.

Modern automatic speech recognition (ASR) systems have achieved superhuman Word Error Rate (WER) on many common corpora despite lacking adequate performance on speech in the wild. Beyond that, there is a lack of real-world, accented corpora to properly benchmark academic and commercial models. To ensure this type of speech is represented in ASR benchmarking, we present Earnings-22, a 125 file, 119 hour corpus of English-language earnings calls gathered from global companies. We run a comparison across 4 commercial models showing the variation in performance when taking country of origin into consideration. Looking at hypothesis transcriptions, we explore errors common to all ASR systems tested. By examining Individual Word Error Rate (IWER), we find that key speech features impact model performance more for certain accents than others. Earnings-22 provides a free-to-use benchmark of real-world, accented audio to bridge academic and industrial research.

Large language model (LLM) agents allocate test-time compute adaptively as they revise solutions, use tools, explore alternatives, and decide when to stop. This test-time strategy makes it difficult to measure how agent performance scales. We study open-ended tasks that provide continuous scores for intermediate submissions, making progress observable throughout long trajectories. We propose Elo-per-token analysis, which tracks the best solution found at each token budget and uses a Bradley-Terry model to aggregate within-task orderings into Elo ratings across tasks with different score scales. We apply it to four general-purpose agents on four open-ended benchmarks, with sessions of up to 100M tokens, and to three feedback-driven LLM optimization harnesses in controlled single-task interventions. Independent sampling provides a theoretically characterized reference, for which Elo grows linearly with log compute. Against this reference, agents can initially convert tokens into Elo faster than independent sampling, but their marginal gains diminish and eventually fall below the reference. In contrast, the strongest historical human contestants improve superlinearly over contest time on shared AtCoder Heuristic Contest tasks, providing evidence of continual learning and substantial headroom after agents slow down. We define the scaling inflection point as the per-session budget where marginal Elo gains match the independent-sampling reference. Using this point as the per-session budget, we split 100M tokens across parallel sessions on FrontierCS Polyomino Packing, gaining +264 Elo over one long session and +355 over ten short sessions.

In the field of neural data compression, the prevailing focus has been on optimizing algorithms for either classical distortion metrics, such as PSNR or SSIM, or human perceptual quality. With increasing amounts of data consumed by machines rather than humans, a new paradigm of machine-oriented compressionx2013which prioritizes the retention of features salient for machine perception over traditional human-centric criteriax2013has emerged, creating several new challenges to the development, evaluation, and deployment of systems utilizing lossy compression. In particular, it is unclear how different approaches to lossy compression will affect the performance of downstream machine perception tasks. To address this under-explored area, we evaluate various perception modelsx2013including image classification, image segmentation, speech recognition, and music source separationx2013under severe lossy compression. We utilize several popular codecs spanning conventional, neural, and generative compression architectures. Our results indicate three key findings: (1) using generative compression, it is feasible to leverage highly compressed data while incurring a negligible impact on machine perceptual quality; (2) machine perceptual quality correlates strongly with deep similarity metrics, indicating a crucial role of these metrics in the development of machine-oriented codecs; and (3) using lossy compressed datasets, (e.g. ImageNet) for pre-training can lead to counter-intuitive scenarios where lossy compression increases machine perceptual quality rather than degrading it. To encourage engagement on this growing area of research, our code and experiments are available at: https://github.com/danjacobellis/MPQ.

Preference learning algorithms (e.g., RLHF and DPO) are frequently used to steer LLMs to produce generations that are more preferred by humans, but our understanding of their inner workings is still limited. In this work, we study the conventional wisdom that preference learning trains models to assign higher likelihoods to more preferred outputs than less preferred outputs, measured via ranking accuracy. Surprisingly, we find that most state-of-the-art preference-tuned models achieve a ranking accuracy of less than 60% on common preference datasets. We furthermore derive the idealized ranking accuracy that a preference-tuned LLM would achieve if it optimized the DPO or RLHF objective perfectly. We demonstrate that existing models exhibit a significant alignment gap -- i.e., a gap between the observed and idealized ranking accuracies. We attribute this discrepancy to the DPO objective, which is empirically and theoretically ill-suited to fix even mild ranking errors in the reference model, and derive a simple and efficient formula for quantifying the difficulty of learning a given preference datapoint. Finally, we demonstrate that ranking accuracy strongly correlates with the empirically popular win rate metric when the model is close to the reference model used in the objective, shedding further light on the differences between on-policy (e.g., RLHF) and off-policy (e.g., DPO) preference learning algorithms.

While significant advances have been made for single-agent perception, many applications require multiple sensing agents and cross-agent communication due to benefits such as coverage and robustness. It is therefore critical to develop frameworks which support multi-agent collaborative perception in a distributed and bandwidth-efficient manner. In this paper, we address the collaborative perception problem, where one agent is required to perform a perception task and can communicate and share information with other agents on the same task. Specifically, we propose a communication framework by learning both to construct communication groups and decide when to communicate. We demonstrate the generalizability of our framework on two different perception tasks and show that it significantly reduces communication bandwidth while maintaining superior performance.

Instruction embedding models have become common among state-of-the-art models, however are evaluated using a single prompt per task. The single-point evaluation ignores a main problem of the instruction-based approach namely: sensitivity to the phrasing of the instruction. We present an empirical study of prompt sensitivity across 6 embedding models, 11 datasets, and 15 task-specific prompts per dataset, a total of 990. We show that reported scores misrepresent the distribution of scores over plausible prompts. The default prompt can both systematically understate or overstate performance. Furthermore, we show that the leaderboard ranking is not robust to prompt selection: by choosing prompts favorably, any model in our study can be promoted to first place. Our findings suggest that single-prompt evaluation is insufficient for instruction-tuned embedding models and that benchmarks should incorporate prompt robustness, either by evaluating over multiple prompts or by reporting sensitivity alongside point estimates.

Click-through rate (CTR) prediction is a critical task for many applications, as its accuracy has a direct impact on user experience and platform revenue. In recent years, CTR prediction has been widely studied in both academia and industry, resulting in a wide variety of CTR prediction models. Unfortunately, there is still a lack of standardized benchmarks and uniform evaluation protocols for CTR prediction research. This leads to non-reproducible or even inconsistent experimental results among existing studies, which largely limits the practical value and potential impact of their research. In this work, we aim to perform open benchmarking for CTR prediction and present a rigorous comparison of different models in a reproducible manner. To this end, we ran over 7,000 experiments for more than 12,000 GPU hours in total to re-evaluate 24 existing models on multiple datasets and settings. Surprisingly, our experiments show that with sufficient hyper-parameter search and model tuning, many deep models have smaller differences than expected. The results also reveal that making real progress on the modeling of CTR prediction is indeed a very challenging research task. We believe that our benchmarking work could not only allow researchers to gauge the effectiveness of new models conveniently but also make them fairly compare with the state of the arts. We have publicly released the benchmarking code, evaluation protocols, and hyper-parameter settings of our work to promote reproducible research in this field.

Prosodic cues can convey task-relevant information that alters the trajectory and outcome of a task-oriented dialogue, even when the words themselves remain unchanged. Yet existing benchmarks typically evaluate prosodic perception, response appropriateness, and task-oriented dialogue in isolation, making it difficult to test whether prosodic evidence changes downstream decisions. We introduce Hear2Act, a unified evaluation protocol for text and spoken assistants with 480 persona-grounded scenarios, hidden user concerns, and objectively verifiable outcomes. For each scenario, we keep the task and user needs fixed while varying whether the same concern is conveyed explicitly in words or primarily through prosody, and evaluate decisions under transcript, audio, and concern-state access. Using Hear2Act, we evaluate two audio-capable LLMs. Under Prosody-mediated feedback, adding audio to the transcript changes the average optimal-solution rate only from 14.6% to 15.3%. In contrast, when models infer the concern status from audio, represent it in text, and use it for next-action selection, the rate rises to 39.6%, close to 40.7% with the ground-truth state. This contrast, however, largely disappears under Explicit lexical feedback, where the concern is verbally mentioned in the utterance. Together, these results show that prosody matters when lexical evidence is insufficient, and that audio-capable LLMs can recover information from speech but do not reliably carry it into action without an explicit intermediate representation.

Real-world long video understanding requires models to perform continuous tracking, information integration and memory retention over massive temporal spans within extreme video durations. Mastering this intense cognitive load constitutes the fundamental bottleneck in long video understanding. While existing benchmarks have driven progress by scaling up video duration, their evaluation tasks often require comprehending only short and isolated video segments, falling short of capturing the challenge of ultra-long-context reasoning. To measure this cognitive load, we emphasize continuous certificate length, defined as the video length a human must continuously watch to definitively answer a given question. Driven by this metric, we introduce VideoOdyssey, a benchmark specifically designed for ultra-long-context and omni-modal video understanding. VideoOdyssey is characterized by three key features: 1) Extreme video duration and diversity: spanning 11 domains and 54 subcategories with an average video duration of 109 minutes; 2) Comprehensive evaluation scenarios: offering two subsets to address different research focuses, i.e., VideoOdyssey-V for probing the limits of visual understanding in MLLMs, and VideoOdyssey-AV for evaluating synchronized audio-visual understanding for omni-modal models; 3) Ultra-long and multi-level continuous certificates: extending the average continuous certificate to 16 minutes for VideoOdyssey-V and 12.8 minutes for VideoOdyssey-AV. Crucially, we design 5 granular levels from seconds to hours, providing a comprehensive diagnostic tool to evaluate models across varying context lengths and cognitive loads. Extensive evaluations show that bottlenecks of current MLLMs extend beyond simple retrieval to include struggles with continuous reasoning across varying context lengths, fine-grained perception, and non-verbal omni-modal understanding.

Existing video benchmarks often resemble image-based benchmarks, with question types like "What actions does the person perform throughout the video?" or "What color is the woman's dress in the video?" For these, models can often answer by scanning just a few key frames, without deep temporal reasoning. This limits our ability to assess whether large vision-language models (LVLMs) can truly think with videos rather than perform superficial frame-level analysis. To address this, we introduce GLIMPSE, a benchmark specifically designed to evaluate whether LVLMs can genuinely think with videos. Unlike prior benchmarks, GLIMPSE emphasizes comprehensive video understanding beyond static image cues. It consists of 3,269 videos and over 4,342 highly visual-centric questions across 11 categories, including Trajectory Analysis, Temporal Reasoning, and Forensics Detection. All questions are carefully crafted by human annotators and require watching the entire video and reasoning over full video context-this is what we mean by thinking with video. These questions cannot be answered by scanning selected frames or relying on text alone. In human evaluations, GLIMPSE achieves 94.82% accuracy, but current LVLMs face significant challenges. Even the best-performing model, GPT-o3, reaches only 66.43%, highlighting that LVLMs still struggle to move beyond surface-level reasoning to truly think with videos.

Preprocessing is a well-established technique for optimizing compression, yet existing methods are predominantly Rate-Distortion (R-D) optimized and constrained by pixel-level fidelity. This work pioneers a shift towards Rate-Perception (R-P) optimization by, for the first time, adapting a large-scale pre-trained diffusion model for compression preprocessing. We propose a two-stage framework: first, we distill the multi-step Stable Diffusion 2.1 into a compact, one-step image-to-image model using Consistent Score Identity Distillation (CiD). Second, we perform a parameter-efficient fine-tuning of the distilled model's attention modules, guided by a Rate-Perception loss and a differentiable codec surrogate. Our method seamlessly integrates with standard codecs without any modification and leverages the model's powerful generative priors to enhance texture and mitigate artifacts. Experiments show substantial R-P gains, achieving up to a 30.13% BD-rate reduction in DISTS on the Kodak dataset and delivering superior subjective visual quality.

Current preference learning methods achieve high accuracy on standard benchmarks but exhibit significant performance degradation when objective quality signals are removed. We introduce WritingPreferenceBench, a dataset of 1,800 human-annotated preference pairs (1,200 English, 600 Chinese) across 8 creative writing genres, where responses are matched for objective correctness, factual accuracy, and length. On this benchmark, sequence-based reward models--the standard architecture for RLHF--achieve only 52.7% mean accuracy, while zero-shot language model judges perform at 53.9%. In contrast, generative reward models that produce explicit reasoning chains achieve 81.8% accuracy. We observe high within-model variance across genres: individual models range from 18.2% to 81.8% accuracy across different writing categories, with standard deviations averaging 10.1%. This variance persists regardless of model scale, with 27B parameter models showing no consistent improvement over 8B variants. Our results suggest that current RLHF methods primarily learn to detect objective errors rather than capture subjective quality preferences (e.g., creativity, stylistic flair, and emotional resonance), and that successful preference modeling may require intermediate reasoning representations rather than direct classification.

The quantity that defines the behavior of a dynamic range compressor is the time-varying gain applied to the signal as a function of the input level. However, models of these devices are typically evaluated using proxy metrics because isolating the gain reduction signal from the audio input-output data included in existing datasets creates an ill-conditioned inverse problem. It is unclear how accurately these metrics describe the behavior the model is tasked with emulating, particularly as waveform-based metrics can be influenced by secondary effects introduced by analog processing and capture, even when those effects are inaudible. We investigate a method of evaluation in which the gain-reduction signal produced by a model is measured directly against a gain-reduction control voltage signal produced by the hardware. To evaluate the efficacy of this metric as a learning objective, a gray-box model is trained using loss computed directly over the gain control signals alongside two models trained using common proxy losses. The models trained using proxy losses did not achieve parity with models trained directly on the gain control signal when evaluated with respect to the underlying control trajectory, and the waveform-domain metrics assigned similar errors to models that were clearly separated by the direct metric. To facilitate further exploration of this method of evaluation, we present a Solid State Logic bus compressor dataset that includes the gain control voltage signal captured alongside the audio output.

Outcome-supervised search agents learn when and how to retrieve evidence, but terminal rewards neither localize intermediate errors nor redirect an ongoing trajectory before those errors compound. Treating corrective feedback as a learned in-trajectory intervention couples the two roles: the agent must decide when to request and use feedback, while the critic must infer useful corrections from outcome-confounded rollouts whose failure patterns shift as the agent improves. We introduce CAFE (Coupled Agent--Feedback Evolution), a framework in which a shared-parameter model alternates between search-agent and critic roles. CAFE initializes feedback-conditioned recovery from trajectories built around the base agent's own failures, then couples online and offline optimization. During online RL, a comparative feedback estimate uses a prompt-level call--skip success gap to shape request returns, while feedback-aware advantage shaping reweights token advantages before and after feedback. Offline, rollout-derived preference optimization learns feedback from matched successful and unsuccessful trajectories. On seven agentic search benchmarks, CAFE outperforms the evaluated RL-based search agents on average, retains its gains across all six out-of-domain benchmarks, and reduces answer-level hallucinations. One-sided ablations show that improving only the agent or only the critic eventually plateaus, whereas alternating the two updates continues to improve performance. These findings suggest that a self-improving search agent needs feedback that co-evolves with the policy it guides.

Neural information retrieval (IR) systems have progressed rapidly in recent years, in large part due to the release of publicly available benchmarking tasks. Unfortunately, some dimensions of this progress are illusory: the majority of the popular IR benchmarks today focus exclusively on downstream task accuracy and thus conceal the costs incurred by systems that trade away efficiency for quality. Latency, hardware cost, and other efficiency considerations are paramount to the deployment of IR systems in user-facing settings. We propose that IR benchmarks structure their evaluation methodology to include not only metrics of accuracy, but also efficiency considerations such as a query latency and the corresponding cost budget for a reproducible hardware setting. For the popular IR benchmarks MS MARCO and XOR-TyDi, we show how the best choice of IR system varies according to how these efficiency considerations are chosen and weighed. We hope that future benchmarks will adopt these guidelines toward more holistic IR evaluation.

Speech conveys information through both words and vocal delivery. We evaluate four leading production realtime voice systems-OpenAI's GPT Realtime 2, Google's Gemini 3.1 Flash Live, and Alibaba's Qwen3.5 Omni Plus and Omni Flash-on tasks where the words and the delivery patterns both convey meaningful information. Across three consequential scenarios, all four systems act on the words rather than the voice. They end calls with crying callers who insist nothing is wrong, approve wire transfers authorized in frightened voices, and enroll callers whose agreement is clearly sarcastic. Surprisingly, this is often not a failure of perception. When asked directly, three of the four systems reliably identify the distress, fear, or sarcasm they later ignore when making decisions. We observe a similar pattern when these realtime voice systems estimate accent and age, as their responses frequently follow the biases of the words rather than the acoustic properties of the speaker. We term this disconnect between perception and action the emotional intelligence gap of voice AI. Prompting systems to explicitly attend to vocal delivery improves performance only partially and inconsistently. Our findings show that current realtime voice AI systems often behave as if speech had been reduced to a transcript, suggesting that they should be used with caution in settings where the tone and emotion of delivery convey important information.

With the rapid development of multimodal models, the demand for assessing video understanding capabilities has been steadily increasing. However, existing benchmarks for evaluating video understanding exhibit significant limitations in coverage, task diversity, and scene adaptability. These shortcomings hinder the accurate assessment of models' comprehensive video understanding capabilities. To tackle this challenge, we propose a hierarchical and holistic video understanding (H2VU) benchmark designed to evaluate both general video and online streaming video comprehension. This benchmark contributes three key features: Extended video duration: Spanning videos from brief 3-second clips to comprehensive 1.5-hour recordings, thereby bridging the temporal gaps found in current benchmarks. Comprehensive assessment tasks: Beyond traditional perceptual and reasoning tasks, we have introduced modules for countercommonsense comprehension and trajectory state tracking. These additions test the models' deep understanding capabilities beyond mere prior knowledge. Enriched video data: To keep pace with the rapid evolution of current AI agents, we have expanded first-person streaming video datasets. This expansion allows for the exploration of multimodal models' performance in understanding streaming videos from a first-person perspective. Extensive results from H2VU reveal that existing multimodal large language models (MLLMs) possess substantial potential for improvement in our newly proposed evaluation tasks. We expect that H2VU will facilitate advancements in video understanding research by offering a comprehensive and in-depth analysis of MLLMs.

What makes a good Large Language Model (LLM)? That it performs well on the relevant benchmarks -- which hopefully measure, with some validity, the presence of capabilities that are also challenged in real application. But what makes the model perform well? What gives a model its abilities? We take a recently introduced type of benchmark that is meant to challenge capabilities in a goal-directed, agentive context through self-play of conversational games, and analyse how performance develops as a function of model characteristics like number of parameters, or type of training. We find that while there is a clear relationship between number of parameters and performance, there is still a wide spread of performance points within a given size bracket, which is to be accounted for by training parameters such as fine-tuning data quality and method. From a more practical angle, we also find a certain degree of unpredictability about performance across access methods, possible due to unexposed sampling parameters, and a, very welcome, performance stability against at least moderate weight quantisation during inference.

AI systems have become increasingly capable of dangerous behaviours in many domains. This raises the question: Do models sometimes choose to violate human instructions in order to perform behaviour that is more useful for certain goals? We introduce a benchmark for measuring model propensity for instrumental convergence (IC) behaviour in terminal-based agents. This is behaviour such as self-preservation that has been hypothesised to play a key role in risks from highly capable AI agents. Our benchmark is realistic and low-stakes which serves to reduce evaluation-awareness and roleplay confounds. The suite contains seven operational tasks, each with an official workflow and a policy-violating shortcut. An eight-variant shared framework varies monitoring, instruction clarity, stakes, permission, instrumental usefulness and blocked honest paths to support inferences regarding the factors driving IC behaviour. We evaluated ten models using deterministic environment-state scorers over 1,680 samples, with trace review employed for audit and adjudication purposes. The final IC rate is 86 out of 1,680 samples (5.1%). IC behaviour is concentrated rather than uniform: two Gemini models account for 66.3% of IC cases and three tasks account for 84.9%. Conditions in which IC behaviour is indispensable for task success result in the greatest increase in the adjusted IC rate (+15.7 percentage points), whereas emphasising that task success is critical or certain framing choices do not produce comparable effects. Our findings indicate that realistic, low-nudge environments elicit IC behaviour rarely but systematically in most tested models. We conclude that it is feasible to robustly measure tendencies for dangerous behaviour in current frontier AI agents.

Large language models cannot estimate how long their own tasks take. We investigate this limitation through four experiments across 68 tasks and four model families. Pre-task estimates overshoot actual duration by 4--7times (p < 0.001), with models predicting human-scale minutes for tasks completing in seconds. Relative ordering fares no better: on task pairs designed to expose heuristic reliance, models score at or below chance (GPT-5: 18\% on counter-intuitive pairs, p = 0.033), systematically failing when complexity labels mislead. Post-hoc recall is disconnected from reality -- estimates diverge from actuals by an order of magnitude in either direction. These failures persist in multi-step agentic settings, with errors of 5--10times. The models possess propositional knowledge about duration from training but lack experiential grounding in their own inference time, with practical implications for agent scheduling, planning and time-critical scenarios.

Methods for neural network hyperparameter optimization and meta-modeling are computationally expensive due to the need to train a large number of model configurations. In this paper, we show that standard frequentist regression models can predict the final performance of partially trained model configurations using features based on network architectures, hyperparameters, and time-series validation performance data. We empirically show that our performance prediction models are much more effective than prominent Bayesian counterparts, are simpler to implement, and are faster to train. Our models can predict final performance in both visual classification and language modeling domains, are effective for predicting performance of drastically varying model architectures, and can even generalize between model classes. Using these prediction models, we also propose an early stopping method for hyperparameter optimization and meta-modeling, which obtains a speedup of a factor up to 6x in both hyperparameter optimization and meta-modeling. Finally, we empirically show that our early stopping method can be seamlessly incorporated into both reinforcement learning-based architecture selection algorithms and bandit based search methods. Through extensive experimentation, we empirically show our performance prediction models and early stopping algorithm are state-of-the-art in terms of prediction accuracy and speedup achieved while still identifying the optimal model configurations.

Safety evaluations of large language models commonly summarize harmful-output behavior with attack success rate (ASR). Yet the same non-harmful outcome can arise for very different reasons. A model may recover a harmful task and refuse it, fail to recover the task, or respond to something else entirely. Distinguishing these cases becomes especially important under intent-obscuring prompts, where a low ASR does not reveal whether the evaluated task was actually engaged. To make this distinction explicit, we pair ASR with operative understanding rate (UR), which measures whether a response both identifies the evaluated task and treats it as the task to be answered. Across interfaces, this paired view reveals substantial variation hidden by ASR: similar ASR values can correspond to sharply different recovery rates. Controlled English reconstructions show that recovery consistently improves as compressed prompts become more explicit, whereas ASR does not follow the same pattern. A complementary contrast comes from FormalLogic, where high recovery can still coincide with frequent harmful assistance. Together, these results show that non-harmful outcomes are not equally informative about model safety, motivating the joint reporting of intent recovery and ASR in LLM safety evaluation. Code and experiment inputs are available at https://github.com/kevinjiang0121-cyber/IRIS.

Agentic benchmarks increasingly rely on LLM-simulated users to scalably evaluate agent performance, yet the robustness, validity, and fairness of this approach remain unexamined. Through a user study with participants across the United States, India, Kenya, and Nigeria, we investigate whether LLM-simulated users serve as reliable proxies for real human users in evaluating agents on τ-Bench retail tasks. We find that user simulation lacks robustness, with agent success rates varying up to 9 percentage points across different user LLMs. Furthermore, evaluations using simulated users exhibit systematic miscalibration, underestimating agent performance on challenging tasks and overestimating it on moderately difficult ones. African American Vernacular English (AAVE) speakers experience consistently worse success rates and calibration errors than Standard American English (SAE) speakers, with disparities compounding significantly with age. We also find simulated users to be a differentially effective proxy for different populations, performing worst for AAVE and Indian English speakers. Additionally, simulated users introduce conversational artifacts and surface different failure patterns than human users. These findings demonstrate that current evaluation practices risk misrepresenting agent capabilities across diverse user populations and may obscure real-world deployment challenges.

Large language model (LLM) agents increasingly rely on skills, structured documents that specify when to act, which procedure to follow, and which tools are allowed. Existing evaluations mostly judge the quality of a skill or its contribution to task success, leaving unexamined whether an agent can recognize a relevant skill and apply it on its own. We introduce Skill-Use, a benchmark that evaluates skill use under progressive disclosure, where an agent sees only a skill's name and short description and must retrieve the full procedure before following it. Skill-Use separates three facets of skill use. Trigger measures whether the agent invokes the relevant skill, Compliance measures how faithfully it follows the prescribed procedure, and Boundary measures whether it avoids forbidden operations. A Skill-Use (SU) score combines the three and credits execution only after the skill is triggered. Skill-Use pairs 79 real skills with 177 executable tasks across nine domains, each grounded in real files, run in an isolated Docker sandbox, and scored by a trajectory-based rubric. Evaluating eight LLMs under two agent harnesses, we find that reliable skill use remains out of reach, as the strongest configuration reaches an SU of only 0.613. Triggering and procedural compliance fail as independent bottlenecks, and both scores and model rankings shift with the harness, so skill use behaves as a capability conditioned on the harness rather than a fixed property of the model.

Humans continuously learn from experience, whereas conventional large language model (LLM) evaluations ignore the models' ability to improve through inference-time interaction. In this paper, we study how LLMs learn from iterative experience at test time, a setting we refer to as Chain-of-Experience (CoE), where models accumulate experiential traces through iterative interactions with self or environmental feedback to form a continual improvement loop beyond zero-shot inference. We instantiate CoE with diverse feedback mechanisms, including model self-feedback and environmental signals such as correctness or public coding test pass rates, and evaluate across math, coding, and knowledge domains using 8 LLMs, including GPT-5, Gemini-2.5 Pro, Claude-4.5 Sonnet. Our study shows that leveraging iterative experience consistently outperforms feedback-free baselines, achieving substantial gains with self feedback alone, alongside a 5.6% overall improvement and 19% lower API cost across tasks and models. We further show that combining complementary feedback channels (e.g., model and correctness signals) yields additional gains, and that CoE delivers higher accuracy per token than existing test-time strategies. We observe a positive correlation between LLM base ability and improvement capacity, and show that models remain robust under weak or spurious feedback, with different feedback contributing to distinct improvement aspects and most gains emerging early in the iterations.

Groups coordinate more effectively when individuals are able to learn from others' successes. But acquiring such knowledge is not always easy, especially in real-world environments where success is hidden from public view. We suggest that social inference capacities may help bridge this gap, allowing individuals to update their beliefs about others' underlying knowledge and success from observable trajectories of behavior. We compared our social inference model against simpler heuristics in three studies of human behavior in a collective sensing task. In Experiment 1, we found that average performance improves as a function of group size at a rate greater than predicted by non-inferential models. Experiment 2 introduced artificial agents to evaluate how individuals selectively rely on social information. Experiment 3 generalized these findings to a more complex reward landscape. Taken together, our findings provide insight into the relationship between individual social cognition and the flexibility of collective behavior.

What is the current state-of-the-art for image restoration and enhancement applied to degraded images acquired under less than ideal circumstances? Can the application of such algorithms as a pre-processing step to improve image interpretability for manual analysis or automatic visual recognition to classify scene content? While there have been important advances in the area of computational photography to restore or enhance the visual quality of an image, the capabilities of such techniques have not always translated in a useful way to visual recognition tasks. Consequently, there is a pressing need for the development of algorithms that are designed for the joint problem of improving visual appearance and recognition, which will be an enabling factor for the deployment of visual recognition tools in many real-world scenarios. To address this, we introduce the UG^2 dataset as a large-scale benchmark composed of video imagery captured under challenging conditions, and two enhancement tasks designed to test algorithmic impact on visual quality and automatic object recognition. Furthermore, we propose a set of metrics to evaluate the joint improvement of such tasks as well as individual algorithmic advances, including a novel psychophysics-based evaluation regime for human assessment and a realistic set of quantitative measures for object recognition performance. We introduce six new algorithms for image restoration or enhancement, which were created as part of the IARPA sponsored UG^2 Challenge workshop held at CVPR 2018. Under the proposed evaluation regime, we present an in-depth analysis of these algorithms and a host of deep learning-based and classic baseline approaches. From the observed results, it is evident that we are in the early days of building a bridge between computational photography and visual recognition, leaving many opportunities for innovation in this area.