Daily Papers

Despite rapid progress in automating scientific research, generating promising and well grounded research solutions remains a central challenge. We isolate research ideation as a standalone task and build our solution on the intuition that a challenge in one field can often be addressed by a mechanism that solved an analogous challenge in another. Accordingly, we introduce IdeaScientist, which decomposes ideation into gap finding, innovation, and report writing, and trains each role with reinforcement learning. These roles identify limitations in related work, draw solution intuitions from analogous problem settings, and develop those intuitions into complete research proposals. To facilitate discovery of insights across domains, we construct the Svalbard Idea Vault, a corpus of 2.77M decomposed research ideas for retrieval, training, and temporally controlled evaluation. Our evaluation restricts access to literature available before a cutoff date and assesses how closely proposed directions align with those later explored in 15K papers authored by human researchers. On Qwen3.6-27B, IdeaScientist outperforms the strongest open-source autoresearch baseline by 14.0%, driven mainly by gains in novelty. On this 27B open backbone, IdeaScientist even outperforms Claude Code SDK with Claude-4.8-Opus and Codex SDK with GPT-5.4, by up to 5.9%.

As large language models (LLMs) reach high scores on established mathematical benchmarks, such as GSM8K and MATH, the research community has turned to International Mathematical Olympiad (IMO) problems to push the evaluation frontier. However, existing Olympiad-level benchmarks suffer from practical constraints that introduce grading noise and potential bias, such as heterogeneous answer formats requiring model-based judges and a reliance on potentially flawed solutions. We introduce RIMO, a two-track benchmark designed to preserve peak Olympiad difficulty while eliminating this evaluation noise. The first track, RIMO-N, rewrites 335 IMO problems to admit a single, unique integer answer, allowing for deterministic correctness checking. The second track, RIMO-P, features 456 proof problems with expert-checked solutions, which are decomposed into a sequence of sub-problems to evaluate the step-by-step reasoning process via an automated grading system. Our benchmarking of ten frontier LLMs, including GPT-4o and Gemini 2.5 Flash, reveals that while these systems excel on older benchmarks, their performance drops sharply on RIMO. These results highlight a substantial gap between current LLM capabilities and actual Olympiad-level reasoning. By providing a challenging yet easy-to-evaluate suite, RIMO offers a high-resolution yardstick for future research, presenting a clear target for closing the profound reasoning gap our findings expose.

Recent agent benchmarks increasingly ground evaluation in executable environments, from code repair to web navigation, app APIs, and function calling. Yet completing consequential work beyond code requires more than producing a plausible response or valid tool call: agents must gather missing information over multiple turns, follow domain policies, coordinate dependent tools, and realize the correct persistent state transition without collateral effects. In this paper, we introduce Thinkingbox, a sandbox for tool-agent-user interaction that provides isolated MCP-compatible tool sessions, complete execution traces, and outcome evaluation over terminal backend state. Built on this sandbox, Thinkingbox-bench contains 507 policy-conditioned workflows across numerous scenarios, including retail, hospitality, auto insurance, neobank internal IT, and consulting IT/HR support. Each attempt is evaluated by task-specific executable checks that accept valid trajectories while rejecting wrong, missing, or extra effects; designated tasks additionally check required properties of the final response. Across proprietary and open-weight models, the strongest achieves 65.36% pass@1, but only 25.25% pass^20. Moreover, many failed trials show clean termination and valid state-changing actions, showing that response or tool-call-level signals are not clear proxies for end-to-end task completion. Thinkingbox-bench reveals a large gap between occasionally finding a successful trajectory and reliably completing stateful business tasks. We release both Thinkingbox and Thinkingbox-Bench: https://github.com/microsoft/thinkingbox

Frontier models have collapsed the cost of writing custom code: a niche problem a specialist sees in their own domain now costs an afternoon. The cost of reviewing and maintaining that code hasn't collapsed. Each solution drifts from the next; understanding one means reading its codebase from scratch. Large enterprises build something centrally governed instead: at worst an off-the-shelf product, at best a graph-orchestration framework wired bespoke per use case, or a low-code platform used as the orchestrator. These are custom every time and limited in scope. Enterprises don't weigh a third option that escapes both constraints: the harness paradigm. Recent work treats the coding-agent harness as enterprise infrastructure rather than a coding tool, converging on three findings: harnesses suffice at the task level and outperform more elaborate architectures on enterprise work (arXiv:2604.00073, arXiv:2604.13107); harness choice accounts for most of the variance in agent benchmark results, more than model choice does (arXiv:2605.23950); and the gap between that finding and enterprise adoption is governance (arXiv:2605.10223, arXiv:2605.18747). We propose an architecture that closes that gap. One harness runs unmodified as the backbone; the code stays identical across every deployment, so reviewing what gets built collapses to reading its instructions file. Section 4 gives four mechanisms: credential-scoped tooling, where each backend gets one generic request tool and a scoped credential instead of a hand-built method; authorization logic outside the harness, so one artifact runs as a cron backbone, a chat-surface engine, and a terminal tool; registration is a side effect of pushing code, collapsing an audit a review of a text file. Built on microcc (<https://pypi.org/project/micro-cc/>), our reference harness.

People overthink; language models over-sample, and the extra effort can talk both into a worse answer. Reasoning systems answer a hard question by sampling it many times (test-time scaling), and the more they draw, the more often a correct answer turns up somewhere, so coverage, the fraction of problems with at least one correct try, climbs and appears to be progress. But a deployed system must return one answer, and choosing it, not knowing which try is right, is selection; selection is capped, and past a point extra samples only make the model surer of a confident mistake, even as every draw adds cost. The gap between climbing coverage and stalled selection, the identifiability gap, is the answer a model can produce but not pick. So the real question is not whether to sample but how far, and the answer is: not far. For picking an answer, the vote has already settled within a few dozen draws, the modal ceiling; for scoring a benchmark, sooner still, the correlation ceiling. Beyond that, extra draws cost compute and add nothing, and can even make the answer worse. This paper turns the cutoff into a single number, the effective number of samples, that any sampling run already reveals. The bottleneck is recognizing a right answer, not generating one.

The performance of quantum annealing for combinatorial optimization is fundamentally limited by the minimum energy gap Δ encountered at quantum phase transitions. We investigate the scaling of Δ with system size N for two paradigmatic quantum spin-glass models: the two-dimensional Edwards-Anderson (2D-EA) and the all-to-all Sherrington-Kirkpatrick (SK) models. Utilizing a newly proposed unbiased energy-gap estimator for continuous-time projection quantum Monte Carlo simulations, complemented by high-performance sparse eigenvalue solvers, we characterize the gap distributions across disorder realizations. It is found that, in the 2D-EA case, the inverse-gap distribution develops a fat tail with infinite variance as N increases. This indicates that the unfavorable super-algebraic scaling of Δ, recently reported for binary couplings [Nature 631, 749 (2024)], persists for the Gaussian disorder considered here, pointing to a universal feature of 2D spin glasses. Conversely, the SK model retains a finite-variance distribution, with the disorder-averaged gap following a rather slow power law, close to Δpropto N^{-1/3}. This finding provides a promising outlook for the potential efficiency of quantum annealers for optimization problems with dense connectivity.

Large Language Models (LLMs) are increasingly deployed in high-stakes clinical applications in India. In many such settings, speakers of Indian languages frequently communicate using romanized text rather than native scripts, yet existing research rarely evaluates this orthographic variation using real-world data. We investigate how romanization impacts the reliability of LLMs in a critical domain: maternal and newborn healthcare triage. We benchmark leading LLMs on a real-world dataset of user-generated queries spanning five Indian languages and Nepali. Our results reveal consistent degradation in performance for romanized messages, with F1 scores trailing those of native scripts by 5-12 points. At our partner maternal health organization in India, this gap could cause nearly 2 million excess errors in triage. Crucially, this performance gap by scripts is not due to a failure in clinical reasoning. We demonstrate that LLMs often correctly infer the semantic intent of romanized queries. Nevertheless, their final classification outputs remain brittle in the presence of orthographic noise in romanized inputs. Our findings highlight a critical safety blind spot in LLM-based health systems: models that appear to understand romanized input may still fail to act on it reliably.

Recent domain generalization (DG) approaches typically use the hypothesis learned on source domains for inference on the unseen target domain. However, such a hypothesis can be arbitrarily far from the optimal one for the target domain, induced by a gap termed ``adaptivity gap''. Without exploiting the domain information from the unseen test samples, adaptivity gap estimation and minimization are intractable, which hinders us to robustify a model to any unknown distribution. In this paper, we first establish a generalization bound that explicitly considers the adaptivity gap. Our bound motivates two strategies to reduce the gap: the first one is ensembling multiple classifiers to enrich the hypothesis space, then we propose effective gap estimation methods for guiding the selection of a better hypothesis for the target. The other method is minimizing the gap directly by adapting model parameters using online target samples. We thus propose Domain-specific Risk Minimization (DRM). During training, DRM models the distributions of different source domains separately; for inference, DRM performs online model steering using the source hypothesis for each arriving target sample. Extensive experiments demonstrate the effectiveness of the proposed DRM for domain generalization with the following advantages: 1) it significantly outperforms competitive baselines on different distributional shift settings; 2) it achieves either comparable or superior accuracies on all source domains compared to vanilla empirical risk minimization; 3) it remains simple and efficient during training, and 4) it is complementary to invariant learning approaches.

Large Language Models (LLMs) have acquired ubiquitous attention for their performances across diverse domains. Our study here searches through LLMs' cognitive abilities and confidence dynamics. We dive deep into understanding the alignment between their self-assessed confidence and actual performance. We exploit these models with diverse sets of questionnaires and real-world scenarios and extract how LLMs exhibit confidence in their responses. Our findings reveal intriguing instances where models demonstrate high confidence even when they answer incorrectly. This is reminiscent of the Dunning-Kruger effect observed in human psychology. In contrast, there are cases where models exhibit low confidence with correct answers revealing potential underestimation biases. Our results underscore the need for a deeper understanding of their cognitive processes. By examining the nuances of LLMs' self-assessment mechanism, this investigation provides noteworthy revelations that serve to advance the functionalities and broaden the potential applications of these formidable language models.

For large language models (LLMs) to be trusted by humans they need to be well-calibrated in the sense that they can accurately assess and communicate how likely it is that their predictions are correct. Recent work has focused on the quality of internal LLM confidence assessments, but the question remains of how well LLMs can communicate this internal model confidence to human users. This paper explores the disparity between external human confidence in an LLM's responses and the internal confidence of the model. Through experiments involving multiple-choice questions, we systematically examine human users' ability to discern the reliability of LLM outputs. Our study focuses on two key areas: (1) assessing users' perception of true LLM confidence and (2) investigating the impact of tailored explanations on this perception. The research highlights that default explanations from LLMs often lead to user overestimation of both the model's confidence and its' accuracy. By modifying the explanations to more accurately reflect the LLM's internal confidence, we observe a significant shift in user perception, aligning it more closely with the model's actual confidence levels. This adjustment in explanatory approach demonstrates potential for enhancing user trust and accuracy in assessing LLM outputs. The findings underscore the importance of transparent communication of confidence levels in LLMs, particularly in high-stakes applications where understanding the reliability of AI-generated information is essential.

Recent research increasingly focuses on training vision-language models (VLMs) with long, detailed image captions. However, small-scale VLMs often struggle to balance the richness of these captions with the risk of hallucinating content during fine-tuning. In this paper, we explore how well VLMs adapt to such captions. To quantify caption quality, we propose Decomposed NLI (DNLI), an evaluation framework that breaks down generated captions into individual propositions, assessing each in isolation. This fine-grained analysis reveals a critical balance between capturing descriptive details and preventing hallucinations. Our findings show that simply reducing caption complexity or employing standard data curation techniques does not effectively resolve this issue. To tackle this challenge, we introduce Knowledge Adapted (KnowAda) fine-tuning, a data-centric approach that automatically adapts training data with the model's existing knowledge and visual understanding. KnowAda minimizes hallucinations while preserving high descriptiveness. We validate this approach across several small-scale VLMs (up to 7B parameters) and dense caption datasets, demonstrating that KnowAda effectively balances hallucination reduction and descriptiveness. Our results show that KnowAda outperforms various baselines in both automatic metrics and human evaluations. We will release our code and models.

Self-improvement is a mechanism in Large Language Model (LLM) pre-training, post-training and test-time inference. We explore a framework where the model verifies its own outputs, filters or reweights data based on this verification, and distills the filtered data. Despite several empirical successes, a fundamental understanding is still lacking. In this work, we initiate a comprehensive, modular and controlled study on LLM self-improvement. We provide a mathematical formulation for self-improvement, which is largely governed by a quantity which we formalize as the generation-verification gap. Through experiments with various model families and tasks, we discover a scaling phenomenon of self-improvement -- a variant of the generation-verification gap scales monotonically with the model pre-training flops. We also examine when self-improvement is possible, an iterative self-improvement procedure, and ways to improve its performance. Our findings not only advance understanding of LLM self-improvement with practical implications, but also open numerous avenues for future research into its capabilities and boundaries.

Large Language Models (LLMs) have shown remarkable performance across various tasks, yet significant disparities remain for non-English languages, and especially native African languages. This paper addresses these disparities by creating approximately 1 million human-translated words of new benchmark data in 8 low-resource African languages, covering a population of over 160 million speakers of: Amharic, Bambara, Igbo, Sepedi (Northern Sotho), Shona, Sesotho (Southern Sotho), Setswana, and Tsonga. Our benchmarks are translations of Winogrande and three sections of MMLU: college medicine, clinical knowledge, and virology. Using the translated benchmarks, we report previously unknown performance gaps between state-of-the-art (SOTA) LLMs in English and African languages. Finally, using results from over 400 fine-tuned models, we explore several methods to reduce the LLM performance gap, including high-quality dataset fine-tuning (using an LLM-as-an-Annotator), cross-lingual transfer, and cultural appropriateness adjustments. Key findings include average mono-lingual improvements of 5.6% with fine-tuning (with 5.4% average mono-lingual improvements when using high-quality data over low-quality data), 2.9% average gains from cross-lingual transfer, and a 3.0% out-of-the-box performance boost on culturally appropriate questions. The publicly available benchmarks, translations, and code from this study support further research and development aimed at creating more inclusive and effective language technologies.

The remarkable success of the autoregressive paradigm has made significant advancement in Multimodal Large Language Models (MLLMs), with powerful models like Show-o, Transfusion and Emu3 achieving notable progress in unified image understanding and generation. For the first time, we uncover a common phenomenon: the understanding capabilities of MLLMs are typically stronger than their generative capabilities, with a significant gap between the two. Building on this insight, we propose HermesFlow, a simple yet general framework designed to seamlessly bridge the gap between understanding and generation in MLLMs. Specifically, we take the homologous data as input to curate homologous preference data of both understanding and generation. Through Pair-DPO and self-play iterative optimization, HermesFlow effectively aligns multimodal understanding and generation using homologous preference data. Extensive experiments demonstrate the significant superiority of our approach over prior methods, particularly in narrowing the gap between multimodal understanding and generation. These findings highlight the potential of HermesFlow as a general alignment framework for next-generation multimodal foundation models. Code: https://github.com/Gen-Verse/HermesFlow

Social biases and belief-driven behaviors can significantly impact Large Language Models (LLMs) decisions on several tasks. As LLMs are increasingly used in multi-agent systems for societal simulations, their ability to model fundamental group psychological characteristics remains critical yet under-explored. In this study, we present a multi-agent framework that simulates belief congruence, a classical group psychology theory that plays a crucial role in shaping societal interactions and preferences. Our findings reveal that LLMs exhibit amplified belief congruence compared to humans, across diverse contexts. We further investigate the implications of this behavior on two downstream tasks: (1) misinformation dissemination and (2) LLM learning, finding that belief congruence in LLMs increases misinformation dissemination and impedes learning. To mitigate these negative impacts, we propose strategies inspired by: (1) contact hypothesis, (2) accuracy nudges, and (3) global citizenship framework. Our results show that the best strategies reduce misinformation dissemination by up to 37% and enhance learning by 11%. Bridging social psychology and AI, our work provides insights to navigate real-world interactions using LLMs while addressing belief-driven biases.

Socioeconomic status (SES) fundamentally influences how people interact with each other and more recently, with digital technologies like Large Language Models (LLMs). While previous research has highlighted the interaction between SES and language technology, it was limited by reliance on proxy metrics and synthetic data. We survey 1,000 individuals from diverse socioeconomic backgrounds about their use of language technologies and generative AI, and collect 6,482 prompts from their previous interactions with LLMs. We find systematic differences across SES groups in language technology usage (i.e., frequency, performed tasks), interaction styles, and topics. Higher SES entails a higher level of abstraction, convey requests more concisely, and topics like 'inclusivity' and 'travel'. Lower SES correlates with higher anthropomorphization of LLMs (using ''hello'' and ''thank you'') and more concrete language. Our findings suggest that while generative language technologies are becoming more accessible to everyone, socioeconomic linguistic differences still stratify their use to exacerbate the digital divide. These differences underscore the importance of considering SES in developing language technologies to accommodate varying linguistic needs rooted in socioeconomic factors and limit the AI Gap across SES groups.

Reinforcement learning with verifiable rewards (RLVR) has enabled large language models (LLMs) to achieve remarkable breakthroughs in reasoning tasks with objective ground-truth answers, such as mathematics and code generation. However, a significant gap remains for non-verifiable tasks, like creative writing and open-ended dialogue, where quality assessment is inherently subjective and lacks definitive references. Existing approaches for these domains often rely on scalar reward models trained with human preferences, which suffer from limited generalization and are prone to reward hacking, such as over-explanation and length bias. In this work, we propose a unified RLVR-based training paradigm that bridges the gap between non-verifiable tasks and verifiable rewards. We introduce a writing-principle-based pairwise Generative Reward Model (GenRM) and a novel Bootstrapped Relative Policy Optimization (BRPO) algorithm. The pairwise writing GenRM leverages self-principled critique to transform subjective assessments into reliable, verifiable rewards, while BRPO enables dynamic, reference-free pairwise comparison by leveraging a bootstrapped response as temporary reference from within group rollouts during RL training. Our approach empowers LLMs to develop robust writing capabilities without supervised fine-tuning, as demonstrated by Writing-Zero, which shows consistent improvement and strong resistance to reward hacking compared to scalar reward baselines. Furthermore, our method achieves competitive results on both in-house and open-source writing benchmarks. Our findings suggest the potential to unify rule-based, reference-based, and reference-free reward modeling under the RLVR framework, thus paving the way for a comprehensive and scalable RL training paradigm applicable across all language tasks.

Large language models have been widely adopted across different tasks, but their auto-regressive generation nature often leads to inefficient resource utilization during inference. While batching is commonly used to increase throughput, performance gains plateau beyond a certain batch size, especially with smaller models, a phenomenon that existing literature typically explains as a shift to the compute-bound regime. In this paper, through an in-depth GPU-level analysis, we reveal that large-batch inference remains memory-bound, with most GPU compute capabilities underutilized due to DRAM bandwidth saturation as the primary bottleneck. To address this, we propose a Batching Configuration Advisor (BCA) that optimizes memory allocation, reducing GPU memory requirements with minimal impact on throughput. The freed memory and underutilized GPU compute capabilities can then be leveraged by concurrent workloads. Specifically, we use model replication to improve serving throughput and GPU utilization. Our findings challenge conventional assumptions about LLM inference, offering new insights and practical strategies for improving resource utilization, particularly for smaller language models.

When evaluating large language models (LLMs) with multiple-choice question answering (MCQA), it is common to end the prompt with the string "Answer:" to facilitate automated answer extraction via next-token probabilities. However, there is no consensus on how to tokenize the space following the colon, often overlooked as a trivial choice. In this paper, we uncover accuracy differences of up to 11% due to this (seemingly irrelevant) tokenization variation as well as reshuffled model rankings, raising concerns about the reliability of LLM comparisons in prior work. Surprisingly, we are able to recommend one specific strategy -- tokenizing the space together with the answer letter -- as we observe consistent and statistically significant performance improvements. Additionally, it improves model calibration, enhancing the reliability of the model's confidence estimates. Our findings underscore the importance of careful evaluation design and highlight the need for standardized, transparent evaluation protocols to ensure reliable and comparable results.

The trajectory of AI development suggests that we will increasingly rely on agent-based systems composed of independently developed agents with different information, privileges, and tools. The success of these systems will critically depend on effective collaboration among these heterogeneous agents, even under partial observability. Despite intense interest, few empirical studies have evaluated such agent-agent collaboration at scale. We propose a collaborative maze-solving benchmark that (i) isolates collaborative capabilities, (ii) modulates problem complexity, (iii) enables scalable automated grading, and (iv) imposes no output-format constraints, preserving ecological plausibility. Using this framework, we evaluate 32 leading open- and closed-source models in solo, homogeneous, and heterogeneous pairings. Our results reveal a "collaboration gap": models that perform well solo often degrade substantially when required to collaborate. Collaboration can break down dramatically; for instance, small distilled models that solve mazes well alone may fail almost completely in certain pairings. We find that starting with the stronger agent often improves outcomes, motivating a "relay inference" approach where the stronger agent leads before handing off to the weaker one, closing much of the gap. Our findings argue for (1) collaboration-aware evaluation, (2) training strategies developed to enhance collaborative capabilities, and (3) interaction design that reliably elicits agents' latent skills, guidance that applies to AI-AI and human-AI collaboration.

Generating realistic synthetic microscopy images is critical for training deep learning models in label-scarce environments, such as cell counting with many cells per image. However, traditional domain adaptation methods often struggle to bridge the domain gap when synthetic images lack the complex textures and visual patterns of real samples. In this work, we adapt the Inversion-Based Style Transfer (InST) framework originally designed for artistic style transfer to biomedical microscopy images. Our method combines latent-space Adaptive Instance Normalization with stochastic inversion in a diffusion model to transfer the style from real fluorescence microscopy images to synthetic ones, while weakly preserving content structure. We evaluate the effectiveness of our InST-based synthetic dataset for downstream cell counting by pre-training and fine-tuning EfficientNet-B0 models on various data sources, including real data, hard-coded synthetic data, and the public Cell200-s dataset. Models trained with our InST-synthesized images achieve up to 37\% lower Mean Absolute Error (MAE) compared to models trained on hard-coded synthetic data, and a 52\% reduction in MAE compared to models trained on Cell200-s (from 53.70 to 25.95 MAE). Notably, our approach also outperforms models trained on real data alone (25.95 vs. 27.74 MAE). Further improvements are achieved when combining InST-synthesized data with lightweight domain adaptation techniques such as DACS with CutMix. These findings demonstrate that InST-based style transfer most effectively reduces the domain gap between synthetic and real microscopy data. Our approach offers a scalable path for enhancing cell counting performance while minimizing manual labeling effort. The source code and resources are publicly available at: https://github.com/MohammadDehghan/InST-Microscopy.

Large Language Models (LLMs) offer a transparent brain with accessible parameters that encode extensive knowledge, which can be analyzed, located and transferred. Consequently, a key research challenge is to transcend traditional knowledge transfer paradigms rooted in symbolic language and achieve genuine Parametric Knowledge Transfer (PKT). Significantly, exploring effective methods for transferring knowledge across LLMs of different scales through parameters presents an intriguing and valuable research direction. In this paper, we first demonstrate Alignment in parametric space is the fundamental prerequisite to achieve successful cross-scale PKT. We redefine the previously explored knowledge transfer as Post-Align PKT (PostPKT), which utilizes extracted parameters for LoRA initialization and requires subsequent fine-tune for alignment. Hence, to reduce cost for further fine-tuning, we introduce a novel Pre-Align PKT (PrePKT) paradigm and propose a solution called LaTen (Locate-Then-Align) that aligns the parametric spaces of LLMs across scales only using several training steps without following training. Comprehensive experiments on four benchmarks demonstrate that both PostPKT and PrePKT face challenges in achieving consistently stable transfer. Through in-depth analysis, we identify Neural Incompatibility as the ethological and parametric structural differences between LLMs of varying scales, presenting fundamental challenges to achieving effective PKT. These findings provide fresh insights into the parametric architectures of LLMs and highlight promising directions for future research on efficient PKT. Our code is available at https://github.com/Trae1ounG/Neural_Incompatibility.

Large Language Models (LLMs) have shown remarkable capabilities in natural language understanding and generation, yet their deployment in enterprise environments reveals a critical limitation: inconsistent adherence to custom instructions. This study presents a comprehensive evaluation of 13 leading LLMs across instruction compliance, response accuracy, and performance metrics in realworld RAG (Retrieval-Augmented Generation) scenarios. Through systematic testing with samples and enterprise-grade evaluation protocols, we demonstrate that instruction following varies dramatically across models, with Claude-Sonnet-4 and GPT-5 achieving the highest results. Our findings reveal the "instruction gap" - a fundamental challenge where models excel at general tasks but struggle with precise instruction adherence required for enterprise deployment. This work provides practical insights for organizations deploying LLM-powered solutions and establishes benchmarks for instruction-following capabilities across major model families.

We introduce DeepSearchQA, a 900-prompt benchmark for evaluating agents on difficult multi-step information-seeking tasks across 17 different fields. Unlike traditional benchmarks that target single answer retrieval or broad-spectrum factuality, DeepSearchQA features a dataset of challenging, handcrafted tasks designed to evaluate an agent's ability to execute complex search plans to generate exhaustive answer lists. This shift in design explicitly tests three critical, yet under-evaluated capabilities: 1) systematic collation of fragmented information from disparate sources, 2) de-duplication and entity resolution to ensure precision, and 3) the ability to reason about stopping criteria within an open-ended search space. Each task is structured as a causal chain, where discovering information for one step is dependent on the successful completion of the previous one, stressing long-horizon planning and context retention. All tasks are grounded in the open web with objectively verifiable answer sets. Our comprehensive evaluation of state-of-the-art agent architectures reveals significant performance limitations: even the most advanced models struggle to balance high recall with precision. We observe distinct failure modes ranging from premature stopping (under-retrieval) to hedging behaviors, where agents cast an overly wide net of low-confidence answers to artificially boost recall. These findings highlight critical headroom in current agent designs and position DeepSearchQA as an essential diagnostic tool for driving future research toward more robust, deep-research capabilities.

Recent advances in unified multimodal models (UMM) have demonstrated remarkable progress in both understanding and generation tasks. However, whether these two capabilities are genuinely aligned and integrated within a single model remains unclear. To investigate this question, we introduce GapEval, a bidirectional benchmark designed to quantify the gap between understanding and generation capabilities, and quantitatively measure the cognitive coherence of the two "unified" directions. Each question can be answered in both modalities (image and text), enabling a symmetric evaluation of a model's bidirectional inference capability and cross-modal consistency. Experiments reveal a persistent gap between the two directions across a wide range of UMMs with different architectures, suggesting that current models achieve only surface-level unification rather than deep cognitive convergence of the two. To further explore the underlying mechanism, we conduct an empirical study from the perspective of knowledge manipulation to illustrate the underlying limitations. Our findings indicate that knowledge within UMMs often remains disjoint. The capability emergence and knowledge across modalities are unsynchronized, paving the way for further exploration.

As NLP evaluation shifts from static benchmarks to multi-turn interactive settings, LLM-based simulators have become widely used as user proxies, serving two roles: generating user turns and providing evaluation signals. Yet, these simulations are frequently assumed to be faithful to real human behaviors, often without rigorous verification. We formalize the Sim2Real gap in user simulation and present the first study running the full τ-bench protocol with real humans (451 participants, 165 tasks), benchmarking 31 LLM simulators across proprietary, open-source, and specialized families using the User-Sim Index (USI), a metric we introduce to quantify how well LLM simulators resemble real user interactive behaviors and feedback. Behaviorally, LLM simulators are excessively cooperative, stylistically uniform, and lack realistic frustration or ambiguity, creating an "easy mode" that inflates agent success rates above the human baseline. In evaluations, real humans provide nuanced judgments across eight quality dimensions while simulated users produce uniformly more positive feedback; rule-based rewards are failing to capture rich feedback signals generated by human users. Overall, higher general model capability does not necessarily yield more faithful user simulation. These findings highlight the importance of human validation when using LLM-based user simulators in the agent development cycle and motivate improved models for user simulation.

Abstract concepts - justice, theory, availability - have no single perceivable referent; in the human brain, their meaning emerges from a web of experiences, affect, and social context. Do large language models (LLMs) ground abstract concepts in a similar way? We study this by replicating property-generation experiments from cognitive science on 21 frontier and open-weight LLMs. Across models and experiments, we find a consistent pattern: when compared to humans, models rely too heavily on word associations, and underproduce properties tied to emotion and internal states. This yields a large and consistent grounding gap: no model exceeds a Pearson correlation r=0.37 with human responses, compared to a human-to-human ceiling above r=0.9. To better interpret this gap, we also replicate a rating experiment on grounding categories and find that here LLMs align more closely with human judgment, and alignment improves as models get larger. We then use sparse autoencoders (SAEs) to inspect whether this information is also reflected in the models' internal features, and we do identify features connected to grounding dimensions such as "sensorimotor" and "social". These findings suggest that current LLMs can recover grounding dimensions when explicitly queried, but do not recruit them in a human-like way when words are generated freely.

Modern language models have historically relied on two dominant design choices: subword tokenization and autoregressive (AR) ordering. These design decisions bake in priors that dictate a model's learning. Recently, two alternative paradigms have challenged this: byte-level modeling, which bypasses static statistically-derived token vocabularies, and masked diffusion modeling (MDM), which conducts parallel, non-sequential generation. Their intersection represents a fully end-to-end modality-agnostic generative prototype; however, removing these structural priors incurs a significant computational cost. In this work, we investigate this cost through a compute-matched scaling study. Our results reveal that the performance penalty of byte modeling is not uniform; across scale, the scaling overhead of byte modeling is worse for MDM than for AR. We hypothesize that this disparity stems from context fragility: while AR's stable causal history allows models to naturally rediscover subword patterns, the MDM objective destroys the local contiguity required to efficiently resolve semantics from raw bytes. Our findings from controlled permutation experiments suggest that future modality-agnostic designs must incorporate alternative structural biases to maintain viable scaling trajectories in the byte regime.

Recent breakthroughs in multi-talker ASR (MT-ASR) and speaker diarization (SD) rely on synthetic data to mitigate the scarcity of large-scale conversational recordings, yet the impact of specific simulation choices remains poorly understood. To mind the gap between simulated mixtures and real-world interactions, we present a study of synthetic data generation for leading MT-ASR (DiCoW) and SD (Sortformer) systems. By introducing FastMSS, a highly efficient open-source simulator, we analyze turn-taking dynamics, source domain, acoustic augmentation, and data mixing strategies. Our findings reveal that optimal simulation recipes are highly task-dependent: increasing speech overlap benefits ASR but degrades diarization. Furthermore, broad source diversity consistently outperforms exact domain matching. Ultimately, synthetic-only training approaches real-data baselines, and combining simulated data with real recordings yields substantial gains over real-only training across both tasks.

Sampling based consistency is widely used for hallucination detection, yet aggregate performance can conceal systematic differences in which errors are detectable. This work studies that heterogeneity across four language models and three factual question answering datasets. Partitioning hallucinations by answer agreement reveals high agreement (Ghost) and low agreement (Flickering) regimes with an apparent detectability gap of 0.35 to 0.46 AUC. Because the statistics used to define the regimes and measure this gap are strongly coupled (|ρ|approx0.94 to 1.00), the raw result is treated as a property of agreement based detection rather than independent evidence. After freezing regime assignments, lexical and semantic response dispersion preserve the asymmetry, with bootstrap 95% intervals excluding zero in all 12 model and dataset settings. A stricter test using individual diffusion trajectories and no cross seed information preserves the asymmetry across all three LLaDA datasets (p<0.005) and directionally across all three Dream datasets, with one reaching significance. The hard regime varies substantially in prevalence across models (16% to 77%), and matched prompts frequently change regimes between models. These findings show that aggregate detection metrics conceal persistent, model dependent heterogeneity in language model failures and motivate regime conditioned evaluation.

Perceived gender inference from names is important for inclusive scientometrics research, yet existing methods perform poorly on non-Western scholars. We address this gap for Chinese names, which appear in two forms, Hanzi and Pinyin, its Romanized form (e.g., Zhang Wei, Li Na). In international bibliographic records, names are often available only in Pinyin. However, a single pinyin name can correspond to multiple distinct hanzi characters with different gender associations, so gender inference from Pinyin alone is intrinsically ambiguous. To address this, we first develop two hanzi models, a regression model that produces a continuous gender score from 0 (male) to 1 (female), and a classification model that outputs one of three labels (male, neutral, female). We train them on a sample of hanzi names from Chinese-language physics publications, rated by 11 native Chinese speakers. On this sample, the models achieve an RMSE of 0.09 and an R^2 of 0.85 (regression), and macro-AUC of 0.97 and a Cohen's κ of 0.81 (classification), demonstrating high performance that is robust to class imbalance. For pinyin names, we propose a pipeline that aggregates hanzi predictions via transliteration and frequency weighting, together with an ambiguity score that flags names whose gender cannot be resolved from Pinyin alone. Benchmarked against three off-the-shelf tools (Namsor, GPT, and Claude), our models achieve the highest agreement with human annotations for both gender scores and gender labels, in both scripts. Applied to an independent sample of Chinese-named authors in APS journals, our models assign a gender to every name, whereas existing Pinyin-based methods leave between 20\% and 100\% of names unclassified. These findings demonstrate the importance of hanzi-level information for reliable gender inference of Chinese names. Code and data are publicly available.

Object tracking has achieved significant progress over the past few years. However, state-of-the-art trackers become increasingly heavy and expensive, which limits their deployments in resource-constrained applications. In this work, we present LightTrack, which uses neural architecture search (NAS) to design more lightweight and efficient object trackers. Comprehensive experiments show that our LightTrack is effective. It can find trackers that achieve superior performance compared to handcrafted SOTA trackers, such as SiamRPN++ and Ocean, while using much fewer model Flops and parameters. Moreover, when deployed on resource-constrained mobile chipsets, the discovered trackers run much faster. For example, on Snapdragon 845 Adreno GPU, LightTrack runs 12times faster than Ocean, while using 13times fewer parameters and 38times fewer Flops. Such improvements might narrow the gap between academic models and industrial deployments in object tracking task. LightTrack is released at https://github.com/researchmm/LightTrack.

Behaviour-Driven Development (BDD) suites accumulate step-text duplication whose maintenance cost is established in prior work. Existing detection techniques require running the tests (Binamungu et al., 2018-2023) or are confined to a single organisation (Irshad et al., 2020-2022), leaving a gap: a purely static, paraphrase-robust, step-level detector usable on any repository. We fill the gap with cukereuse, an open-source Python CLI combining exact hashing, Levenshtein ratio, and sentence-transformer embeddings in a layered pipeline, released alongside an empirical corpus of 347 public GitHub repositories, 23,667 parsed .feature files, and 1,113,616 Gherkin steps. The step-weighted exact-duplicate rate is 80.2 %; the median-repository rate is 58.6 % (Spearman rho = 0.51 with size). The top hybrid cluster groups 20.7k occurrences across 2.2k files. Against 1,020 pairs manually labelled by the three authors under a released rubric (inter-annotator Fleiss' kappa = 0.84 on a 60-pair overlap), we report precision, recall, and F1 with bootstrap 95 % CIs under two protocols: the primary rubric and a score-free second-pass relabelling. The strongest honest pair-level number is near-exact at F1 = 0.822 on score-free labels; the primary-rubric semantic F1 = 0.906 is inflated by a stratification artefact that pins recall at 1.000. Lexical baselines (SourcererCC-style, NiCad-style) reach primary F1 = 0.761 and 0.799. The paper also presents a CDN-structured critique of Gherkin (Cognitive Dimensions of Notations); eight of fourteen dimensions are rated problematic or unsupported. The tool, corpus, labelled pairs, rubric, and pipeline are released under permissive licences.

Though diffusion-based video generation has witnessed rapid progress, the inference results of existing models still exhibit unsatisfactory temporal consistency and unnatural dynamics. In this paper, we delve deep into the noise initialization of video diffusion models, and discover an implicit training-inference gap that attributes to the unsatisfactory inference quality. Our key findings are: 1) the spatial-temporal frequency distribution of the initial latent at inference is intrinsically different from that for training, and 2) the denoising process is significantly influenced by the low-frequency components of the initial noise. Motivated by these observations, we propose a concise yet effective inference sampling strategy, FreeInit, which significantly improves temporal consistency of videos generated by diffusion models. Through iteratively refining the spatial-temporal low-frequency components of the initial latent during inference, FreeInit is able to compensate the initialization gap between training and inference, thus effectively improving the subject appearance and temporal consistency of generation results. Extensive experiments demonstrate that FreeInit consistently enhances the generation results of various text-to-video generation models without additional training.

Contrastive audio-language pretraining yields powerful joint representations, yet a persistent audio-text modality gap limits the benefits of coupling multimodal encoders with large language models (LLMs). We present Diffusion-Link, a diffusion-based modality-bridging module that generatively maps audio embeddings into the text-embedding distribution. The module is trained at the output embedding from the frozen multimodal encoder and implemented as a lightweight network with three residual MLP blocks. To assess the effect of Diffusion-Link on multimodal encoder-LLM coupling, we evaluate on Automatic Audio Captioning (AAC); to our knowledge, this is the first application of diffusion-based modality bridging to AAC. We report two results. (1) Modality-gap analysis: on similarity and geometric criteria, Diffusion-Link reduces the modality gap the most among prior diffusion-based methods and shows a collective migration of audio embeddings toward the text distribution. (2) Downstream AAC: attaching Diffusion-Link to the same multimodal LLM baseline achieves state-of-the-art on AudioCaps in both zero-shot and fully supervised captioning without external knowledge, with relative gains up to 52.5% and 7.5%, respectively. These findings show that closing the modality gap is pivotal for effective coupling between multimodal encoders and LLMs, and diffusion-based modality bridging offers a promising direction beyond knowledge-retrieval-centric designs. Code will be released upon acceptance https://github.com/DevKiHyun/Diffusion-Link

We present a comprehensive evaluation of the ability of large language models (LLMs) to process culturally grounded language, specifically to understand and pragmatically use figurative expressions that encode local knowledge and cultural nuance. Using figurative language as a proxy for cultural nuance and local knowledge, we design evaluation tasks for contextual understanding, pragmatic use, and connotation interpretation in Arabic and English. We evaluate 22 open- and closed-source LLMs on Egyptian Arabic idioms, multidialectal Arabic proverbs, and English proverbs. Our results show a consistent hierarchy: the average accuracy for Arabic proverbs is 4.29% lower than for English proverbs, and performance for Egyptian idioms is 10.28% lower than for Arabic proverbs. For the pragmatic use task, accuracy drops by 14.07% relative to understanding, though providing contextual idiomatic sentences improves accuracy by 10.66%. Models also struggle with connotative meaning, reaching at most 85.58% agreement with human annotators on idioms with 100% inter-annotator agreement. These findings demonstrate that figurative language serves as an effective diagnostic for cultural reasoning: while LLMs can often interpret figurative meaning, they face challenges in using it appropriately. To support future research, we release Kinayat, the first dataset of Egyptian Arabic idioms designed for both figurative understanding and pragmatic use evaluation.

Multimodal learning has recently gained significant popularity, demonstrating impressive performance across various zero-shot classification tasks and a range of perceptive and generative applications. Models such as Contrastive Language-Image Pretraining (CLIP) are designed to bridge different modalities, such as images and text, by learning a shared representation space through contrastive learning. Despite their success, the working mechanisms underlying multimodal learning are not yet well understood. Notably, these models often exhibit a modality gap, where different modalities occupy distinct regions within the shared representation space. In this work, we conduct an in-depth analysis of the emergence of modality gap by characterizing the gradient flow learning dynamics. Specifically, we identify the critical roles of mismatched data pairs and a learnable temperature parameter in causing and perpetuating the modality gap during training. Furthermore, our theoretical insights are validated through experiments on practical CLIP models. These findings provide principled guidance for mitigating the modality gap, including strategies such as appropriate temperature scheduling and modality swapping. Additionally, we demonstrate that closing the modality gap leads to improved performance on tasks such as image-text retrieval.

Vision-language models (VLMs) lag behind text-only language models on mathematical reasoning when the same problems are presented as images rather than text. We empirically characterize this as a modality gap: the same question in text form yields markedly higher accuracy than its visually typeset counterpart, due to compounded failures in reading dense formulas, layout, and mixed symbolic-diagrammatic context. First, we introduce VisTIRA (Vision and Tool-Integrated Reasoning Agent), a tool-integrated reasoning framework that enables structured problem solving by iteratively decomposing a given math problem (as an image) into natural language rationales and executable Python steps to determine the final answer. Second, we build a framework to measure and improve visual math reasoning: a LaTeX-based pipeline that converts chain-of-thought math corpora (e.g., NuminaMath) into challenging image counterparts, and a large set of synthetic tool-use trajectories derived from a real-world, homework-style image dataset (called SnapAsk) for fine-tuning VLMs. Our experiments show that tool-integrated supervision improves image-based reasoning, and OCR grounding can further narrow the gap for smaller models, although its benefit diminishes at scale. These findings highlight that modality gap severity inversely correlates with model size, and that structured reasoning and OCR-based grounding are complementary strategies for advancing visual mathematical reasoning.

Stylistic personalization - making LLMs write in a specific individual's style, rather than merely adapting to task preferences - lacks evaluation grounded in authorship science. We show that grounding evaluation in authorship verification theory transforms what benchmarks can measure. Drawing on three measurement traditions - LUAR, a trained authorship verification model; an LLM-as-judge with decoupled trait matching; and classical function-word stylometrics - we evaluate four inference-time personalization methods across 50 authors and 1,000 generations. The theory-grounded metric, LUAR, provides what ad hoc alternatives cannot: calibrated baselines, with a human ceiling of 0.756 and a cross-author floor of 0.626, that give scores absolute meaning. All methods score below this floor, from 0.484 to 0.508, exposing an authorship gap invisible to uncalibrated metrics. The three metrics produce near-zero pairwise correlations, with absolute r less than 0.07, confirming that without theoretical grounding, metric choice determines conclusions: an LLM judge declares a clear winner while LUAR finds no meaningful differentiation. These findings demonstrate the theory-benchmark cycle in action: authorship theory exposes evaluation failures that ad hoc benchmarks miss.

Multimodal LLMs are increasingly used to assist scientific peer review, where a core requirement is verifying whether claims in a paper are supported by its evidence. Prior work has shown that models perform substantially better at this task when the evidence is a table than when it is a chart of the same underlying data. This raises the question of whether models fail to extract information from charts, or do they extract it but fail to use it when forming their prediction? We study this question through layer-wise linear probing and attention analysis on three open-weight VLMs over table and chart evidence, representing the same underlying data. We find consistent evidence for the latter. Chart information is encoded in the models' intermediate representations but does not reach the prediction position, a gap that is absent for tables and holds across all conditions tested. Attention analysis further reveals that this disconnect takes two architecturally distinct forms across model families. These findings reframe the table-chart gap as a failure of how encoded visual information is routed at prediction time, rather than a failure of encoding itself.

Instruction following is central to language-conditioned robot policies: language should determine what to do when the same scene permits multiple valid actions. Yet successful execution alone cannot establish whether a policy follows the instruction or infers the task from the scene. We study this ambiguity through scene-preserving instruction interventions, using valid target substitutions, arbitrary nouns, and unrelated sentences while holding the scene fixed. We evaluate vision-language-action (VLA) policies and world-action models (WAMs) in simulation and in real-world experiments. Our analysis addresses three questions: (a) Does task success imply instruction following? When instructions request a different visible object, all evaluated policies predominantly approach and pick up the incorrect original target associated with the scene. (b) Is this failure caused by language insensitivity? Instruction perturbations affect task performance. A layerwise action lens shows intermediate action predictions respond to these perturbations. Linear probes accurately recover instructed targets, indicating modified instructions are encoded despite rarely determining target selection. (c) Why does encoded language fail to control action? Attention analysis indicates weak instruction-token contributions to action generation. Target-token attention can remain focused on the original object, revealing a mismatch between target encoding and visual grounding. UMAP and shared non-negative matrix factorization show target information remains accessible within representations increasingly organized by scene identity. Our findings expose a grounding gap concealed by nominal success and provide a diagnostic framework. They further establish a concrete criterion for progress: policies should reliably follow valid changes in user intent, even when they conflict with scene-favored behavior.

Accurate 6D pose estimation for robotic harvesting is fundamentally hindered by the biological deformability and high intra-class shape variability of agricultural produce. Instance-level methods fail in this setting, as obtaining exact 3D models for every unique piece of produce is practically infeasible, while category-level approaches that rely on a fixed template suffer significant accuracy degradation when the prior deviates from the true instance geometry. To bridge such lack of robustness to deformation, we introduce PEAR (Pose and dEformation of Agricultural pRoduce), the first benchmark providing joint 6D pose and per-instance 3D deformation ground truth across 8 produce categories, acquired via a robotic manipulator for high annotation accuracy. Using PEAR, we show that state-of-the-art methods suffer up to 6x performance degradation when faced with the inherent geometric deviations of real-world produce. Motivated by this finding, we propose SEED (Simultaneous Estimation of posE and Deformation), a unified RGB-only framework that jointly predicts 6D pose and explicit lattice deformations from a single image across multiple produce categories. Trained entirely on synthetic data with generative texture augmentation applied at the UV level, SEED outperforms MegaPose on 6 out of 8 categories under identical RGB-only conditions, demonstrating that explicit shape modeling is a critical step toward reliable pose estimation in agricultural robotics.

Person retrieval in surveillance videos often depends on attributes described by witnesses or operators. However, the most useful cues in practice are not always common appearance descriptions (e.g., gender, clothing color), but rare and distinctive attributes that can sharply reduce the search space (e.g., holding a weapon, lying on the ground). Existing text-based person retrieval benchmarks and methods largely focus on identity-centric retrieval with common pedestrian descriptions, leaving such retrieval-critical attributes underexplored. In this paper, we introduce Open-Attribute Person Retrieval (OAPR), a practical retrieval setting that aims to retrieve all pedestrian instances matching a given attribute query, including rare or previously unseen visual concepts, regardless of identity. To support this task, we construct EPAD, an Expanded Pedestrian Attribute Dataset with 267,885 pedestrian images and a unified vocabulary of 65 attributes, including safety-critical actions, assistive devices, and object interactions that are rarely covered in prior benchmarks. We further propose GAP-CLIP, a lightweight CLIP-based framework that learns gated attribute-aware body-part representations for OAPR. Extensive experiments on EPAD demonstrate that GAP-CLIP achieves the strongest top-K retrieval performance on the full attribute space and on out-of-distribution attributes. The code and dataset are available at https://github.com/mlnjeongpark/Open-Attribute-Person-Retrieval.

Due to the expensive costs of collecting labels in multi-label classification datasets, partially annotated multi-label classification has become an emerging field in computer vision. One baseline approach to this task is to assume unobserved labels as negative labels, but this assumption induces label noise as a form of false negative. To understand the negative impact caused by false negative labels, we study how these labels affect the model's explanation. We observe that the explanation of two models, trained with full and partial labels each, highlights similar regions but with different scaling, where the latter tends to have lower attribution scores. Based on these findings, we propose to boost the attribution scores of the model trained with partial labels to make its explanation resemble that of the model trained with full labels. Even with the conceptually simple approach, the multi-label classification performance improves by a large margin in three different datasets on a single positive label setting and one on a large-scale partial label setting. Code is available at https://github.com/youngwk/BridgeGapExplanationPAMC.

Scaling dense retrievers to larger large language model (LLM) backbones has been a dominant strategy for improving their retrieval effectiveness. However, this has substantial cost implications: larger backbones require more expensive hardware (e.g. GPUs with more memory) and lead to higher indexing and querying costs (latency, energy consumption). In this paper, we challenge this paradigm by introducing PromptPRF, a feature-based pseudo-relevance feedback (PRF) framework that enables small LLM-based dense retrievers to achieve effectiveness comparable to much larger models. PromptPRF uses LLMs to extract query-independent, structured and unstructured features (e.g., entities, summaries, chain-of-thought keywords, essay) from top-ranked documents. These features are generated offline and integrated into dense query representations via prompting, enabling efficient retrieval without additional training. Unlike prior methods such as GRF, which rely on online, query-specific generation and sparse retrieval, PromptPRF decouples feedback generation from query processing and supports dense retrievers in a fully zero-shot setting. Experiments on TREC DL and BEIR benchmarks demonstrate that PromptPRF consistently improves retrieval effectiveness and offers favourable cost-effectiveness trade-offs. We further present ablation studies to understand the role of positional feedback and analyse the interplay between feature extractor size, PRF depth, and model performance. Our findings demonstrate that with effective PRF design, scaling the retriever is not always necessary, narrowing the gap between small and large models while reducing inference cost.

Recent Vision-and-Language Navigation (VLN) advancements are promising, but their idealized assumptions about robot movement and control fail to reflect physically embodied deployment challenges. To bridge this gap, we introduce VLN-PE, a physically realistic VLN platform supporting humanoid, quadruped, and wheeled robots. For the first time, we systematically evaluate several ego-centric VLN methods in physical robotic settings across different technical pipelines, including classification models for single-step discrete action prediction, a diffusion model for dense waypoint prediction, and a train-free, map-based large language model (LLM) integrated with path planning. Our results reveal significant performance degradation due to limited robot observation space, environmental lighting variations, and physical challenges like collisions and falls. This also exposes locomotion constraints for legged robots in complex environments. VLN-PE is highly extensible, allowing seamless integration of new scenes beyond MP3D, thereby enabling more comprehensive VLN evaluation. Despite the weak generalization of current models in physical deployment, VLN-PE provides a new pathway for improving cross-embodiment's overall adaptability. We hope our findings and tools inspire the community to rethink VLN limitations and advance robust, practical VLN models. The code is available at https://crystalsixone.github.io/vln_pe.github.io/.

Large Language Models (LLMs) and other automated techniques have been increasingly used to support software developers by generating software artifacts such as code snippets, patches, and comments. However, accurately assessing the correctness of these generated artifacts remains a significant challenge. On one hand, human evaluation provides high accuracy but is labor-intensive and lacks scalability. On the other hand, many automatic evaluation metrics are scalable and require minimal human effort, but they often fail to accurately reflect the actual correctness of generated software artifacts. In this paper, we present SE-Jury, the first evaluation metric for LLM-as-Ensemble-Judge specifically designed to accurately assess the correctness of generated software artifacts. SE-Jury first defines five distinct evaluation strategies, each implemented by an independent judge. A dynamic team selection mechanism then identifies the most appropriate subset of judges as a team to produce a final correctness score through ensembling. We evaluate SE-Jury across a diverse set of software engineering (SE) benchmarks that span three popular SE tasks: code generation, automated program repair, and code summarization. Results demonstrate that SE-Jury consistently achieves a higher correlation with human judgments, with improvements ranging from 29.6% to 140.8% over existing automatic metrics. SE-Jury reaches agreement levels with human annotators that are close to inter-annotator agreement in code generation and program repair. These findings underscore SE-Jury's potential as a scalable and reliable alternative to human evaluation in these SE tasks.

Materials combining both a high refractive index and a wide band gap are of great interest for optoelectronic and sensor applications. However, these two properties are typically described by an inverse correlation with high refractive index appearing in small gap materials and vice-versa. Here, we conduct a first-principles high-throughput study on more than 4000 semiconductors (with a special focus on oxides). Our data confirm the general inverse trend between refractive index and band gap but interesting outliers are also identified. The data are then analyzed through a simple model involving two main descriptors: the average optical gap and the effective frequency. The former can be determined directly from the electronic structure of the compounds, but the latter cannot. This calls for further analysis in order to obtain a predictive model. Nonetheless, it turns out that the negative effect of a large band gap on the refractive index can counterbalanced in two ways: (i) by limiting the difference between the direct band gap and the average optical gap which can be realized by a narrow distribution in energy of the optical transitions and (ii) by increasing the effective frequency which can be achieved through either a high number of transitions from the top of the valence band to the bottom of the conduction or a high average probability for these transitions. Focusing on oxides, we use our data to investigate how the chemistry influences this inverse relationship and rationalize why certain classes of materials would perform better. Our findings can be used to search for new compounds in many optical applications both in the linear and non-linear regime (waveguides, optical modulators, laser, frequency converter, etc.).

Are tool-calling LLM agents equally safe throughout a conversation? We discover they are not: agents are most vulnerable at the very start of a session and become substantially safer after a few regular agentic tasks -- a phenomenon we term the cold-start safety gap. To study this systematically, we introduce Safety Over Depth for Agents (SODA), a benchmark that controls how many regular agentic tasks the agent completes before encountering a safety threat, supporting up to 20 preceding tasks. Evaluating 7 models from 4 families, safety improves by 9--52% as the number of preceding regular agentic tasks increases from zero to twenty. Representation analysis confirms that model hidden states gradually shift toward a safety-aligned region as more preceding tasks are present. By systematically studying which part of the preceding conversation matters most, we find that the regular agentic tasks themselves are the primary driver of safety, while the agent's own prior responses have less effect on safety but are essential for preserving later utility. This conclusion is further supported by evaluation on open-source safety benchmarks (AgentHarm, Agent Safety Bench) and utility benchmarks (BFCL, API-Bank), confirming that warming up the agent with regular agentic tasks before deployment makes it safer and preserves full capability. Based on these findings, we recommend a simple deployment strategy: having the agent complete a few regular agentic tasks before possible exposure to safety-critical requests mitigates the cold-start safety gap. Our code is available at https://github.com/Trustworthy-ML-Lab/Agent-Cold-Start-Safety-Gap

Studies of human reasoning have shown that people are typically stronger at evaluating reasoning than producing it from scratch. In contrast, large reasoning models (LRMs) are trained to excel at producing long chains of reasoning to solve complex problems. How then do LRMs perform at evaluating reasons? We investigate this with the Valid-Answer-Invalid-Reasoning (VAIR) dataset: math problems and solutions with trivial reasoning flaws but valid answers, designed to isolate reasoning evaluation from the confound of reasoning production. Unlike humans, who we find are only 6% worse at grading than solving such problems, we find a substantial production-evaluation gap in LRMs: frontier models score as low as 48% when evaluating VAIR solutions, despite near-perfect solution production. Why this enigma? Through chain-of-thought (CoT) analysis, we find evidence of an answer confirmation bias: LRMs often produce then check for the correct answer instead of carefully verifying each step, fabricating rationalizations even when noticing anomalous reasoning. Linear probes corroborate this, showing that while LRM activations encode some representation of valid reasoning, they fail to robustly represent VAIR solutions as invalid. Causal patching of the final answer's representations causes LRM verdicts and activations to flip, demonstrating that answer validity is responsible for models' confirmation biases. These findings indicate an outstanding limitation in dominant approaches to reasoning training, which incentivize LRMs to produce and confirm reasoning towards correct answers, but not to robustly evaluate the underlying reasons.

A central challenge in developing Multimodal Large Language Models (MLLMs) is effectively integrating heterogeneous inputs into a cohesive reasoning engine. Current paradigms predominantly rely on modular architectures that introduce modality-specific encoders and cross-modal fusion mechanisms. However, these designs are fundamentally bottlenecked by a geometric modality gap, forcing the LLM to expend significant computational capacity on geometric reconciliation rather than deep cross-modal reasoning. In this work, we formally characterize this modality gap and theoretically demonstrate that native architectures, specifically those employing a unified vocabulary, intrinsically maintain a zero-gap state across all hidden layers. Guided by these theoretical findings, we propose One Tokenizer, a native architecture that maps all modalities directly into a shared token space. We empirically validate this framework on a DNA--text multimodal testbed. Our extensive evaluations reveal that by achieving seamless integration within the LLM's native latent space, One Tokenizer consistently outperforms encoder-based modular counterparts, providing a fundamentally superior framework for deep biological reasoning.

Generative UI tools promise to democratize UI design by turning natural language descriptions into complete interfaces. Alongside the interface, these tools generate user-facing design rationales that explain their layout, accessibility, and design choices. However, it remains unclear whether these stated rationales are actually reflected in the interfaces they produce. We call this disconnect ``Design Theater'': plausible and confident design rationales that have little relationship to the actual implementation. To study this phenomenon, we introduce a benchmark and three metrics for measuring Design Theater. The benchmark includes 24 UI generation tasks spanning structural, styling, and functional design requirements. Using this benchmark, we evaluate 120 interfaces created by five generative UI tools. On average, over 25\% of user-facing design rationales are not implemented in the generated interface, and the implementation failure increases to 34\% for functional requirements. Tools recognize roughly half of the UX principles embedded in prompts (mean = 0.54), with four of five tools implementing 6\% or fewer functional principles. We also measure interface similarity across tools and find convergence in visual appearance and layout organization, with greater variation in color choices. Overall, we contribute: 1) the concept of Design Theater; 2) a benchmark with metrics for assessing whether the stated reasoning of generative UI tools is reflected in their implementations; 3) and findings from a systematic evaluation of these tools. We discuss what these findings mean for the design and evaluation of generative UI tools.

Activation steering has become a widely used approach for controlling language models during explicit chain-of-thought (CoT) reasoning, motivating its extension to latent CoT. However, we find that steering continuous thoughts produces substantially weaker effects on subsequent language generation than steering explicit CoT, even when the hidden representations are moved by comparable amounts. We first show that task information remains identifiable in continuous thoughts. Hence, we hypothesize a latent-to-language transition gap, in which an intervention effect in latent space fails to transfer to language generation. Two further results support this hypothesis: the output distribution changes abruptly at the transition boundary, and task-related directions exert much weaker bidirectional control in latent CoT than in explicit CoT. These findings identify the transition interface as a central target for evaluating and designing future latent-steering methods.

Spatial reasoning is central to navigation and robotics, yet measuring model capabilities on these tasks remains difficult. Existing benchmarks evaluate models in a one-shot setting, requiring full solution generation in a single response, unlike humans, who work in interactive environments step-by-step. We introduce Spatial-Gym, a Gymnasium environment that isolates spatial constraint reasoning by testing pathfinding in 2D-grid puzzles as a sequential decision task with optional backtracking. We evaluate eight models in three settings (one-shot, step-by-step, step-by-step with backtracking) against human, random, and A* baselines on 500 episodes. The best model, GPT-OSS 120B, achieves a solve rate of 16.0%, 82 points below the human baseline (98.0%). Step-by-step format helps weaker models (up to +5.4%) by removing formatting errors, but hurts stronger models (up to 5.6%) by constraining global planning. Backtracking improves episode completion, but increases solve rate only for weaker models; stronger models rarely backtrack and do not benefit from it. Our experiments have three key findings: (1) models fail to scale reasoning effort with difficulty, (2) vision models receiving images of the spatial environment reduce solve rate by 73%, and (3) extended chain-of-thought reasoning retains a 3-5x accuracy advantage over standard inference even in the step-by-step setting. Spatial-Gym enables diagnosis of model limitations and provides a framework for improving spatial reasoning through reinforcement learning.

Referring Image Segmentation is a comprehensive task to segment an object referred by a textual query from an image. In nature, the level of difficulty in this task is affected by the existence of similar objects and the complexity of the referring expression. Recent RIS models still show a significant performance gap between easy and hard scenarios. We pose that the bottleneck exists in the data, and propose a simple but powerful data augmentation method, Negative-mined Mosaic Augmentation (NeMo). This method augments a training image into a mosaic with three other negative images carefully curated by a pretrained multimodal alignment model, e.g., CLIP, to make the sample more challenging. We discover that it is critical to properly adjust the difficulty level, neither too ambiguous nor too trivial. The augmented training data encourages the RIS model to recognize subtle differences and relationships between similar visual entities and to concretely understand the whole expression to locate the right target better. Our approach shows consistent improvements on various datasets and models, verified by extensive experiments.

Multilingual text-to-image (T2I) models have advanced rapidly in terms of visual realism and semantic alignment, and are now widely utilized. Yet outputs vary across cultural contexts: because language carries cultural connotations, images synthesized from multilingual prompts should preserve cross-lingual cultural consistency. We conduct a comprehensive analysis showing that current T2I models often produce culturally neutral or English-biased results under multilingual prompts. Analyses of two representative models indicate that the issue stems not from missing cultural knowledge but from insufficient activation of culture-related representations. We propose a probing method that localizes culture-sensitive signals to a small set of neurons in a few fixed layers. Guided by this finding, we introduce two complementary alignment strategies: (1) inference-time cultural activation that amplifies the identified neurons without backbone fine-tuned; and (2) layer-targeted cultural enhancement that updates only culturally relevant layers. Experiments on our CultureBench demonstrate consistent improvements over strong baselines in cultural consistency while preserving fidelity and diversity.

Recent advances in transformer-based architectures which are pre-trained in self-supervised manner have shown great promise in several machine learning tasks. In the audio domain, such architectures have also been successfully utilised in the field of speech emotion recognition (SER). However, existing works have not evaluated the influence of model size and pre-training data on downstream performance, and have shown limited attention to generalisation, robustness, fairness, and efficiency. The present contribution conducts a thorough analysis of these aspects on several pre-trained variants of wav2vec 2.0 and HuBERT that we fine-tuned on the dimensions arousal, dominance, and valence of MSP-Podcast, while additionally using IEMOCAP and MOSI to test cross-corpus generalisation. To the best of our knowledge, we obtain the top performance for valence prediction without use of explicit linguistic information, with a concordance correlation coefficient (CCC) of .638 on MSP-Podcast. Furthermore, our investigations reveal that transformer-based architectures are more robust to small perturbations compared to a CNN-based baseline and fair with respect to biological sex groups, but not towards individual speakers. Finally, we are the first to show that their extraordinary success on valence is based on implicit linguistic information learnt during fine-tuning of the transformer layers, which explains why they perform on-par with recent multimodal approaches that explicitly utilise textual information. Our findings collectively paint the following picture: transformer-based architectures constitute the new state-of-the-art in SER, but further advances are needed to mitigate remaining robustness and individual speaker issues. To make our findings reproducible, we release the best performing model to the community.

Clinical deployment of automated brain MRI analysis faces a fundamental challenge: clinical data is heterogeneous and noisy, and high-quality labels are prohibitively costly to obtain. Self-supervised learning (SSL) can address this by leveraging the vast amounts of unlabeled data produced in clinical workflows to train robust foundation models that adapt out-of-domain with minimal supervision. However, the development of foundation models for brain MRI has been limited by small pretraining datasets and in-domain benchmarking focused on high-quality, research-grade data. To address this gap, we organized the FOMO25 challenge as a satellite event at MICCAI 2025. FOMO25 provided participants with a large pretraining dataset, FOMO60K, and evaluated models on data sourced directly from clinical workflows in few-shot and out-of-domain settings. Tasks covered infarct classification, meningioma segmentation, and brain age regression, and considered both models trained on FOMO60K (method track) and any data (open track). Nineteen foundation models from sixteen teams were evaluated using a standardized containerized pipeline. Results show that (a) self-supervised pretraining improves generalization on clinical data under domain shift, with the strongest models trained out-of-domain surpassing supervised baselines trained in-domain. (b) No single pretraining objective benefits all tasks: MAE favors segmentation, hybrid reconstruction-contrastive objectives favor classification, and (c) strong performance was achieved by small pretrained models, and improvements from scaling model size and training duration did not yield reliable benefits.

Post-training quantization is often treated as a semantically neutral optimization for edge deployment of Large Language Models. When a full-precision source checkpoint is evaluated and quantization is applied downstream without equivalent re-evaluation, this workflow creates a structural validation--deployment gap: because quantization is a many-to-one mapping over parameter space, source-precision certification does not guarantee behavioral equivalence in the deployed configuration. We formalize this gap through Quantization Behavioral Equivalence Classes (QBECs) and prove that QBEC membership does not imply behavioral equivalence, providing a theoretical basis for quantization-triggered backdoor attacks. Building on a three-stage adversarial fine-tuning framework, we embed latent malicious payloads into models that satisfy the source-precision checks used in our evaluation, yet activate targeted adversarial behavior upon INT8 or 4-bit compression. We evaluate this threat in two operationally motivated scenarios, tactical machine translation and political content analysis, extending prior work from decoder-only causal LMs to multilingual encoder-decoder sequence-to-sequence models. Results show that backdoored translation models move from zero measured friend--foe corruption at repaired FP16 to up to 85.02% inversion after quantization, and that a paired stance classifier measures an ideological shift of up to ΔBias=0.33 upon compression. A cross-quantizer transferability analysis further shows that attack persistence varies across quantization schemes and model architectures, rather than being determined by nominal bit-width alone. These findings demonstrate that source-precision auditing alone does not rule out quantization-triggered behavior and that the final deployed configuration must be included in behavioral certification for trustworthy edge AI.

Mechanistic interpretability reverse-engineers transformer circuits one input at a time, leaving observed mechanisms without guarantees over bounded input neighbourhoods. We address this gap with a framework based on constrained polynomial-zonotope (CPZ) propagation that lifts mechanistic-interpretability observations from a single input to certified statements over a bounded set of perturbations. Three internal-attention queries (top-k stability, evidence mass, and attention entropy) are formulated as tractable programs over the simplex of attention weights, and CPZ propagation through transformer blocks is shown to preserve the softmax simplex and the LayerNorm zero-mean identity exactly. A recursive Jacobian zonotope construction extends the same certificates across layer depth by linearising the block stack at the input and avoids per-layer generator growth. We instantiate the framework on transformer attention; the resulting certificates offer a way to sharpen mechanistic statements that single-input inspection cannot resolve on its own, and to inform downstream decisions in regimes where empirical heuristics may be misleading.

Inspired by the great success of language model (LM)-based pre-training, recent studies in visual document understanding have explored LM-based pre-training methods for modeling text within document images. Among them, pre-training that reads all text from an image has shown promise, but often exhibits instability and even fails when applied to broader domains, such as those involving both visual documents and scene text images. This is a substantial limitation for real-world scenarios, where the processing of text image inputs in diverse domains is essential. In this paper, we investigate effective pre-training tasks in the broader domains and also propose a novel pre-training method called SCOB that leverages character-wise supervised contrastive learning with online text rendering to effectively pre-train document and scene text domains by bridging the domain gap. Moreover, SCOB enables weakly supervised learning, significantly reducing annotation costs. Extensive benchmarks demonstrate that SCOB generally improves vanilla pre-training methods and achieves comparable performance to state-of-the-art methods. Our findings suggest that SCOB can be served generally and effectively for read-type pre-training methods. The code will be available at https://github.com/naver-ai/scob.

The BabyLM Challenge is a community effort to close the data-efficiency gap between human and computational language learners. Participants compete to optimize language model training on a fixed language data budget of 100 million words or less. This year, we released improved text corpora, as well as a vision-and-language corpus to facilitate research into cognitively plausible vision language models. Submissions were compared on evaluation tasks targeting grammatical ability, (visual) question answering, pragmatic abilities, and grounding, among other abilities. Participants could submit to a 10M-word text-only track, a 100M-word text-only track, and/or a 100M-word and image multimodal track. From 31 submissions employing diverse methods, a hybrid causal-masked language model architecture outperformed other approaches. No submissions outperformed the baselines in the multimodal track. In follow-up analyses, we found a strong relationship between training FLOPs and average performance across tasks, and that the best-performing submissions proposed changes to the training data, training objective, and model architecture. This year's BabyLM Challenge shows that there is still significant room for innovation in this setting, in particular for image-text modeling, but community-driven research can yield actionable insights about effective strategies for small-scale language modeling.

We present ReXGroundingCT, the first publicly available dataset to link free-text radiology findings with pixel-level segmentations in 3D chest CT scans that is manually annotated. While prior datasets have relied on structured labels or predefined categories, ReXGroundingCT captures the full expressiveness of clinical language represented in free text and grounds it to spatially localized 3D segmentation annotations in volumetric imaging. This addresses a critical gap in medical AI: the ability to connect complex, descriptive text, such as "3 mm nodule in the left lower lobe", to its precise anatomical location in three-dimensional space, a capability essential for grounded radiology report generation systems. The dataset comprises 3,142 non-contrast chest CT scans paired with standardized radiology reports from the CT-RATE dataset. Using a systematic three-stage pipeline, GPT-4 was used to extract positive lung and pleural findings, which were then manually segmented by expert annotators. A total of 8,028 findings across 16,301 entities were annotated, with quality control performed by board-certified radiologists. Approximately 79% of findings are focal abnormalities, while 21% are non-focal. The training set includes up to three representative segmentations per finding, while the validation and test sets contain exhaustive labels for each finding entity. ReXGroundingCT establishes a new benchmark for developing and evaluating sentence-level grounding and free-text medical segmentation models in chest CT. The dataset can be accessed at https://huggingface.co/datasets/rajpurkarlab/ReXGroundingCT.

Large Language Models (LLMs) have significantly advanced medical question-answering by leveraging extensive clinical data and medical literature. However, the rapid evolution of medical knowledge and the labor-intensive process of manually updating domain-specific resources pose challenges to the reliability of these systems. To address this, we introduce Agentic Medical Graph-RAG (AMG-RAG), a comprehensive framework that automates the construction and continuous updating of medical knowledge graphs, integrates reasoning, and retrieves current external evidence, such as PubMed and WikiSearch. By dynamically linking new findings and complex medical concepts, AMG-RAG not only improves accuracy but also enhances interpretability in medical queries. Evaluations on the MEDQA and MEDMCQA benchmarks demonstrate the effectiveness of AMG-RAG, achieving an F1 score of 74.1 percent on MEDQA and an accuracy of 66.34 percent on MEDMCQA, outperforming both comparable models and those 10 to 100 times larger. Notably, these improvements are achieved without increasing computational overhead, highlighting the critical role of automated knowledge graph generation and external evidence retrieval in delivering up-to-date, trustworthy medical insights.

Large language models (LLMs) are increasingly deployed in privacy-critical and personalization-oriented scenarios, yet the role of context length in shaping privacy leakage and personalization effectiveness remains largely unexplored. We introduce a large-scale benchmark, PAPerBench, to systematically study how increasing context length influences both personalization quality and privacy protection in LLMs. The benchmark comprises approximately 29,000 instances with context lengths ranging from 1K to 256K tokens, yielding a total of 377K evaluation questions. It jointly evaluates personalization performance and privacy risks across diverse scenarios, enabling controlled analysis of long-context model behavior. Extensive evaluations across state-of-the-art LLMs reveal consistent performance degradation in both personalization and privacy as context length increases. We further provide a theoretical analysis of attention dilution under context scaling, explaining this behavior as an inherent limitation of soft attention in fixed-capacity Transformers. The empirical and theoretical findings together suggest a general scaling gap in current models -- long context, less focus. We release the benchmark to support reproducible evaluation and future research on scalable privacy and personalization. Code and data are available at https://github.com/SafeRL-Lab/PAPerBench

Reasoning in vision-language models (VLMs) has recently attracted significant attention due to its broad applicability across diverse downstream tasks. However, it remains unclear whether the superior performance of VLMs stems from genuine vision-grounded reasoning or relies predominantly on the reasoning capabilities of their textual backbones. To systematically measure this, we introduce CrossMath, a novel multimodal reasoning benchmark designed for controlled cross-modal comparisons. Specifically, we construct each problem in text-only, image-only, and image+text formats guaranteeing identical task-relevant information, verified by human annotators. This rigorous alignment effectively isolates modality-specific reasoning differences while eliminating confounding factors such as information mismatch. Extensive evaluation of state-of-the-art VLMs reveals a consistent phenomenon: a substantial performance gap between textual and visual reasoning. Notably, VLMs excel with text-only inputs, whereas incorporating visual data (image+text) frequently degrades performance compared to the text-only baseline. These findings indicate that current VLMs conduct reasoning primarily in the textual space, with limited genuine reliance on visual evidence. To mitigate this limitation, we curate a CrossMath training set for VLM fine-tuning. Empirical evaluations demonstrate that fine-tuning on this training set significantly boosts reasoning performance across all individual and joint modalities, while yielding robust gains on two general visual reasoning tasks. Source code is available at https://github.com/xuyige/CrossMath.

Background: Deep learning models are typically trained using stochastic gradient descent or one of its variants. These methods update the weights using their gradient, estimated from a small fraction of the training data. It has been observed that when using large batch sizes there is a persistent degradation in generalization performance - known as the "generalization gap" phenomena. Identifying the origin of this gap and closing it had remained an open problem. Contributions: We examine the initial high learning rate training phase. We find that the weight distance from its initialization grows logarithmically with the number of weight updates. We therefore propose a "random walk on random landscape" statistical model which is known to exhibit similar "ultra-slow" diffusion behavior. Following this hypothesis we conducted experiments to show empirically that the "generalization gap" stems from the relatively small number of updates rather than the batch size, and can be completely eliminated by adapting the training regime used. We further investigate different techniques to train models in the large-batch regime and present a novel algorithm named "Ghost Batch Normalization" which enables significant decrease in the generalization gap without increasing the number of updates. To validate our findings we conduct several additional experiments on MNIST, CIFAR-10, CIFAR-100 and ImageNet. Finally, we reassess common practices and beliefs concerning training of deep models and suggest they may not be optimal to achieve good generalization.

Step-level reward models (SRMs) can significantly enhance mathematical reasoning performance through process supervision or step-level preference alignment based on reinforcement learning. The performance of SRMs is pivotal, as they serve as critical guidelines, ensuring that each step in the reasoning process is aligned with desired outcomes. Recently, AlphaZero-like methods, where Monte Carlo Tree Search (MCTS) is employed for automatic step-level preference annotation, have proven particularly effective. However, the precise mechanisms behind the success of SRMs remain largely unexplored. To address this gap, this study delves into the counterintuitive aspects of SRMs, particularly focusing on MCTS-based approaches. Our findings reveal that the removal of natural language descriptions of thought processes has minimal impact on the efficacy of SRMs. Furthermore, we demonstrate that SRMs are adept at assessing the complex logical coherence present in mathematical language while having difficulty in natural language. These insights provide a nuanced understanding of the core elements that drive effective step-level reward modeling in mathematical reasoning. By shedding light on these mechanisms, this study offers valuable guidance for developing more efficient and streamlined SRMs, which can be achieved by focusing on the crucial parts of mathematical reasoning.

Large language models (LLMs) are renowned for their extensive linguistic knowledge and strong generalization capabilities, but their high computational demands make them unsuitable for resource-constrained environments. In contrast, small language models (SLMs) are computationally efficient but often lack the broad generalization capacity of LLMs. To bridge this gap, we propose PiFi, a novel framework that combines the strengths of both LLMs and SLMs to achieve high performance while maintaining efficiency. PiFi integrates a single frozen layer from an LLM into a SLM and fine-tunes the combined model for specific tasks, boosting performance without a significant increase in computational cost. We show that PiFi delivers consistent performance improvements across a range of natural language processing tasks, including both natural language understanding and generation. Moreover, our findings demonstrate PiFi's ability to effectively leverage LLM knowledge, enhancing generalization to unseen domains and facilitating the transfer of linguistic abilities.

Multimodal large language models (MLLMs) can process text presented as images, yet they often perform worse than when the same content is provided as textual tokens. We systematically diagnose this "modality gap" by evaluating seven MLLMs across seven benchmarks in five input modes, spanning both synthetically rendered text and realistic document images from arXiv PDFs to Wikipedia pages. We find that the modality gap is task- and data-dependent. For example, math tasks degrade by over 60 points on synthetic renderings, while natural document images often match or exceed text-mode performance. Rendering choices such as font and resolution are strong confounds, with font alone swinging accuracy by up to 47 percentage points. To understand this, we conduct a grounded-theory error analysis of over 4,000 examples, revealing that image mode selectively amplifies reading errors (calculation and formatting failures) while leaving knowledge and reasoning errors largely unchanged, and that some models exhibit a chain-of-thought reasoning collapse under visual input. Motivated by these findings, we propose a self-distillation method that trains the model on its own pure text reasoning traces paired with image inputs, raising image-mode accuracy on GSM8K from 30.71% to 92.72% and transferring to unseen benchmarks without catastrophic forgetting. Overall, our study provides a systematic understanding of the modality gap and suggests a practical path toward improving visual text understanding in multimodal language models.

Embodied navigation is a fundamental capability of embodied intelligence, enabling robots to move and interact within physical environments. However, existing navigation tasks primarily focus on predefined object navigation or instruction following, which significantly differs from human needs in real-world scenarios involving complex, open-ended scenes. To bridge this gap, we introduce a challenging long-horizon navigation task that requires understanding high-level human instructions and performing spatial-aware object navigation in real-world environments. Existing embodied navigation methods struggle with such tasks due to their limitations in comprehending high-level human instructions and localizing objects with an open vocabulary. In this paper, we propose NavA^3, a hierarchical framework divided into two stages: global and local policies. In the global policy, we leverage the reasoning capabilities of Reasoning-VLM to parse high-level human instructions and integrate them with global 3D scene views. This allows us to reason and navigate to regions most likely to contain the goal object. In the local policy, we have collected a dataset of 1.0 million samples of spatial-aware object affordances to train the NaviAfford model (PointingVLM), which provides robust open-vocabulary object localization and spatial awareness for precise goal identification and navigation in complex environments. Extensive experiments demonstrate that NavA^3 achieves SOTA results in navigation performance and can successfully complete longhorizon navigation tasks across different robot embodiments in real-world settings, paving the way for universal embodied navigation. The dataset and code will be made available. Project website: https://NavigationA3.github.io/.

We present the Conversational Data Retrieval (CDR) benchmark, the first comprehensive test set for evaluating systems that retrieve conversation data for product insights. With 1.6k queries across five analytical tasks and 9.1k conversations, our benchmark provides a reliable standard for measuring conversational data retrieval performance. Our evaluation of 16 popular embedding models shows that even the best models reach only around NDCG@10 of 0.51, revealing a substantial gap between document and conversational data retrieval capabilities. Our work identifies unique challenges in conversational data retrieval (implicit state recognition, turn dynamics, contextual references) while providing practical query templates and detailed error analysis across different task categories. The benchmark dataset and code are available at https://github.com/l-yohai/CDR-Benchmark.

Although modern neural networks often generalize to new combinations of familiar concepts, the conditions that enable such compositionality have long been an open question. In this work, we study the systematicity gap in visual question answering: the performance difference between reasoning on previously seen and unseen combinations of object attributes. To test, we introduce a novel diagnostic dataset, CLEVR-HOPE. We find that the systematicity gap is not reduced by increasing the quantity of training data, but is reduced by increasing the diversity of training data. In particular, our experiments suggest that the more distinct attribute type combinations are seen during training, the more systematic we can expect the resulting model to be.

Niche-domain Indic ASR -- digit strings, currency amounts, addresses, brand names, English/Indic codemix -- is under-served by both open-source SOTA and commercial systems. On a synthesised entity-dense Telugu test set (held-out by synthesis system), vasista22/whisper-telugu-large-v2 (open SOTA) achieves Entity-Hit-Rate (EHR) 0.027 and Deepgram Nova-3 (commercial) 0.16. We close this gap with a self-contained TTS<->STT flywheel: an open-source Indic TTS pipeline synthesises ~22,000 entity-dense Indic-English code-mix utterances at <$50 marginal cost, and a LoRA fine-tune on top of vasista22 achieves EHR 0.473 on the held-out test (17x over open SOTA, 3x over commercial), with read-prose regression bounded to +6.6 pp WER on FLEURS-Te. Cross-language: beta-Hi 0.337 (7x vs vasista22) and beta-Ta 0.543 (22x vs vasista22, 22x vs Deepgram); on Hindi where Deepgram has substantial entity coverage, the flywheel underperforms commercial. All three beta models fall below pre-registered EHR targets (0.75 for Te, 0.65 for Hi/Ta); we report honestly. A native-human-recorded sanity check (n=20 Telugu) confirms transfer to real speech (beta-Te EHR 0.516 on native vs 0.473 on synth). An EDSA-isolation ablation (LoRA on FLEURS-Te alone) yields EHR 0.020 on the same held-out, attributing ~100% of the gain to the EDSA corpus. We additionally report a language-conditional finding: vanilla Whisper-large-v3 has Telugu-specific Script Collapse (SFR 0.46-0.71) that a per-language LoRA corrects (SFR 0.81-0.97), but the recipe is contraindicated on Hindi and Tamil where vanilla SFR >= 0.98. Code, holdouts, predictions, EDSA corpus, and entity dictionaries are released open-source.

Rapidly-exploring random trees (RRTs) have been widely adopted for robot motion planning due to their robustness and theoretical guarantees. However, existing RRT-based planners require explicit goal configurations specified as numerical joint angles, while many practical applications provide goal specifications through visual observations such as images or demonstration videos where precise goal configurations are unavailable. In this paper, we propose visual-RRT (vRRT), a motion planner that enables visual-goal planning by unifying gradient-based exploitation from differentiable robot rendering with sampling-based exploration from RRTs. We further introduce (i) a frontier-based exploration-exploitation strategy that adaptively prioritizes visually promising search regions, and (ii) inertial gradient tree expansion that inherits optimization states across tree branches for momentum-consistent gradient exploitation. Extensive experiments across various robot manipulators including Franka, UR5e, and Fetch demonstrate that vRRT achieves effective visual-goal planning in both simulated and real-world settings, bridging the gap between sampling-based planning and vision-centric robot applications. Our code is available at https://sgvr.kaist.ac.kr/Visual-RRT.

Large language models (LLMs) demonstrate exceptional performance on complex reasoning tasks. However, despite their strong reasoning capabilities in high-resource languages (e.g., English and Chinese), a significant performance gap persists in other languages. To investigate this gap in Korean, we introduce HRM8K, a benchmark comprising 8,011 English-Korean parallel bilingual math problems. Through systematic analysis of model behaviors, we identify a key finding: these performance disparities stem primarily from difficulties in comprehending non-English inputs, rather than limitations in reasoning capabilities. Based on these findings, we propose UST (Understand, Solve, and Translate), a method that strategically uses English as an anchor for reasoning and solution generation. By fine-tuning the model on 130k synthetically generated data points, UST achieves a 10.91% improvement on the HRM8K benchmark and reduces the multilingual performance gap from 11.6% to 0.7%. Additionally, we show that improvements from UST generalize effectively to different Korean domains, demonstrating that capabilities acquired from machine-verifiable content can be generalized to other areas. We publicly release the benchmark, training dataset, and models.

Pre-trained multi-modal Vision-Language Models like CLIP are widely used off-the-shelf for a variety of applications. In this paper, we show that the common practice of individually exploiting the text or image encoders of these powerful multi-modal models is highly suboptimal for intra-modal tasks like image-to-image retrieval. We argue that this is inherently due to the CLIP-style inter-modal contrastive loss that does not enforce any intra-modal constraints, leading to what we call intra-modal misalignment. To demonstrate this, we leverage two optimization-based modality inversion techniques that map representations from their input modality to the complementary one without any need for auxiliary data or additional trained adapters. We empirically show that, in the intra-modal tasks of image-to-image and text-to-text retrieval, approaching these tasks inter-modally significantly improves performance with respect to intra-modal baselines on more than fifteen datasets. Additionally, we demonstrate that approaching a native inter-modal task (e.g. zero-shot image classification) intra-modally decreases performance, further validating our findings. Finally, we show that incorporating an intra-modal term in the pre-training objective or narrowing the modality gap between the text and image feature embedding spaces helps reduce the intra-modal misalignment. The code is publicly available at: https://github.com/miccunifi/Cross-the-Gap.

We present our shared task on text-based emotion detection, covering more than 30 languages from seven distinct language families. These languages are predominantly low-resource and spoken across various continents. The data instances are multi-labeled into six emotional classes, with additional datasets in 11 languages annotated for emotion intensity. Participants were asked to predict labels in three tracks: (a) emotion labels in monolingual settings, (b) emotion intensity scores, and (c) emotion labels in cross-lingual settings. The task attracted over 700 participants. We received final submissions from more than 200 teams and 93 system description papers. We report baseline results, as well as findings on the best-performing systems, the most common approaches, and the most effective methods across various tracks and languages. The datasets for this task are publicly available.

Recent large vision-language models (LVLMs) have been applied to diverse VQA tasks. However, achieving practical performance typically requires task-specific fine-tuning with large numbers of image-text pairs, which are costly to collect. In this work, we study text-centric training, a setting where only textual descriptions are available and no real images are provided, as a paradigm for low-cost data scaling. Unlike images, whose collection is often restricted by privacy constraints and scarcity in niche domains, text is widely available. Moreover, text is easily editable, enabling automatic diversification and expansion with LLMs at minimal human effort. While this offers clear advantages over image collection in terms of scalability and cost, training on raw text without images still yields limited gains on VQA tasks because of the image-text modality gap. To address this issue, we propose a Text-Printed Image (TPI), which generates synthetic images by directly rendering the given textual description on a plain white canvas. This simple rendering projects text into the image modality and can be integrated into arbitrary existing LVLM training pipelines at low cost. Moreover, TPI preserves the semantics of the text, whereas text-to-image models often fail to do. Across four models and seven benchmarks, our systematic experiments show that TPI enables more effective text-centric training than synthetic images generated by a diffusion model. We further explore TPI as a low-cost data-augmentation strategy and demonstrate its practical utility. Overall, our findings highlight the significant potential of text-centric training and, more broadly, chart a path toward fully automated data generation for LVLMs.

A benchmark score means nothing without knowing what a trivial method achieves and what the best possible method could achieve. We construct both bounds for a task with a rare kind of ground truth: predicting which sentences a crowd of readers -- highlighting for their own purposes, unpaid, uninstructed, and blind to each other -- marked in 120 web documents. The floor is naive truncation (lead); the ceiling is a split-half oracle: half the crowd predicting the other half. The gap between them is +0.2028 AP [+0.1698, +0.2342, domain-clustered], and three findings structure it. First, the gap is semantic: position and length features recover 5% of it. Second, frontier language models reach 35-53% of it zero-shot -- far above classical baselines, far below the crowd; a state-of-the-art prompt compressor (LLMLingua-2) lands below the floor, indistinguishable from random selection. Third, an unweighted cross-vendor fusion of five frontier rankings plus a position prior reaches 60%, beating the best single model by +0.0159 [+0.0044, +0.0269; Holm p=0.019] -- a gain that survives ablation of its best member, split-half arm selection, prompt paraphrase, and label, gate, and seed perturbations, and was CONFIRMED by a pre-registered replication on 217 independent documents (+0.0179, Holm p=0.042). Finally, the bracket compresses: distilling the fusion into one open-weight 8B student that reads the whole document retains 90% of the fusion's edge and reaches statistical parity with the strongest single frontier model (+0.0070 [-0.0068, +0.0200]), where a local-context student retains only 63% -- the crowd's signal lives in document-level structure, and the cheapest known improvement is to ask several different models and average.

The rapid development of Multimodal Large Language Models (MLLMs) has expanded their capabilities from image comprehension to video understanding. However, most of these MLLMs focus primarily on offline video comprehension, necessitating extensive processing of all video frames before any queries can be made. This presents a significant gap compared to the human ability to watch, listen, think, and respond to streaming inputs in real time, highlighting the limitations of current MLLMs. In this paper, we introduce StreamingBench, the first comprehensive benchmark designed to evaluate the streaming video understanding capabilities of MLLMs. StreamingBench assesses three core aspects of streaming video understanding: (1) real-time visual understanding, (2) omni-source understanding, and (3) contextual understanding. The benchmark consists of 18 tasks, featuring 900 videos and 4,500 human-curated QA pairs. Each video features five questions presented at different time points to simulate a continuous streaming scenario. We conduct experiments on StreamingBench with 13 open-source and proprietary MLLMs and find that even the most advanced proprietary MLLMs like Gemini 1.5 Pro and GPT-4o perform significantly below human-level streaming video understanding capabilities. We hope our work can facilitate further advancements for MLLMs, empowering them to approach human-level video comprehension and interaction in more realistic scenarios.

Volatiles like H_2O are present as ice in solids in the outer cold regions of protoplanetary disks and as vapor in the warm inner regions within the water snow line. Icy pebbles drifting inwards from the outer disk sublimate after crossing the snow line, enriching the inner disk with solid mass and water vapor. Meanwhile, proto-planets forming within the disk open gaps in the disk gas, creating traps against the inward drift of pebbles and in turn reducing water enrichment in the inner disk. Recent disk observations from millimeter interferometry and infrared spectroscopy have supported this broad picture by finding a correlation between the outer radial distribution of pebbles and the properties of inner water vapor spectra. In this work, we aim at further informing previous and future observations by building on previous models to explore pebble drift in disks with multiple gaps. We systematically explore multiple gap locations and their depths (equivalent to specific masses of planets forming within), and different particle sizes to study their impact on inner disk water enrichment. We find that the presence of close-in deep gaps carved by a Jupiter-mass planet is likely crucial for blocking icy pebble delivery into the inner disk, while planets with lower masses only provide leaky traps. We also find that disks with multiple gaps show lower vapor enrichment in the inner disk. Altogether, these model results support the idea that inner disk water delivery and planet formation are regulated by the mass and location of the most massive planets.

Large language models (LLMs) tend to verbalize confidence scores that are largely detached from their actual accuracy, yet the geometric relationship governing this behavior remain poorly understood. In this work, we present a mechanistic interpretability analysis of verbalized confidence, using linear probes and contrastive activation addition (CAA) steering to show that calibration and verbalized confidence signals are encoded linearly but are orthogonal to one another -- a finding consistent across three open-weight models and four datasets. Interestingly, when models are prompted to simultaneously reason through a problem and verbalize a confidence score, the reasoning process disrupts the verbalized confidence direction, exacerbating miscalibration. We term this the "Reasoning Contamination Effect." Leveraging this insight, we introduce a two-stage adaptive steering pipeline that reads the model's internal accuracy estimate and steers verbalized output to match it, substantially improving calibration alignment across all evaluated models.

Recent advances in monocular depth estimation have been made by incorporating natural language as additional guidance. Although yielding impressive results, the impact of the language prior, particularly in terms of generalization and robustness, remains unexplored. In this paper, we address this gap by quantifying the impact of this prior and introduce methods to benchmark its effectiveness across various settings. We generate "low-level" sentences that convey object-centric, three-dimensional spatial relationships, incorporate them as additional language priors and evaluate their downstream impact on depth estimation. Our key finding is that current language-guided depth estimators perform optimally only with scene-level descriptions and counter-intuitively fare worse with low level descriptions. Despite leveraging additional data, these methods are not robust to directed adversarial attacks and decline in performance with an increase in distribution shift. Finally, to provide a foundation for future research, we identify points of failures and offer insights to better understand these shortcomings. With an increasing number of methods using language for depth estimation, our findings highlight the opportunities and pitfalls that require careful consideration for effective deployment in real-world settings

Recent studies empirically indicate that language models (LMs) encode rich world knowledge beyond mere semantics, attracting significant attention across various fields. However, in the recommendation domain, it remains uncertain whether LMs implicitly encode user preference information. Contrary to prevailing understanding that LMs and traditional recommenders learn two distinct representation spaces due to the huge gap in language and behavior modeling objectives, this work re-examines such understanding and explores extracting a recommendation space directly from the language representation space. Surprisingly, our findings demonstrate that item representations, when linearly mapped from advanced LM representations, yield superior recommendation performance. This outcome suggests the possible homomorphism between the advanced language representation space and an effective item representation space for recommendation, implying that collaborative signals may be implicitly encoded within LMs. Motivated by these findings, we explore the possibility of designing advanced collaborative filtering (CF) models purely based on language representations without ID-based embeddings. To be specific, we incorporate several crucial components to build a simple yet effective model, with item titles as the input. Empirical results show that such a simple model can outperform leading ID-based CF models, which sheds light on using language representations for better recommendation. Moreover, we systematically analyze this simple model and find several key features for using advanced language representations: a good initialization for item representations, zero-shot recommendation abilities, and being aware of user intention. Our findings highlight the connection between language modeling and behavior modeling, which can inspire both natural language processing and recommender system communities.

Large language models increasingly understand dialectal English, yet still produce only standard, US-leaning English, leaving dialectal generation, the harder half of the problem, largely unaddressed. We introduce DiaLLM, which continually pretrains three open-weight language model families on the International Corpus of English and applies implicit and explicit post-training paradigms, each combined with three model alignment strategies, giving the first controlled comparison of these components across Australian, Indian, and Northern British English. Our results reveal that dialectal robustness and generation are dissociated: benchmarks are shaped by continual pretraining and SFT, while alignment visibly reshapes generation in ways benchmarks do not capture. Explicit variety-targeted adaptation produces output reliably recognised as dialectal and preferred over broad alignment, yet the method that most aggressively optimises the dialectal reward is not preferred by human evaluators. Independent linguistic analysis corroborates this reward-quality gap, most clearly on two of the three families. No single alignment method dominates, and closing the gap will require richer reward designs and continued investment in dialectal resources. We release all code, checkpoints, and preference datasets.

Large language model (LLM) agents are increasingly used across the scientific research lifecycle: ideation, literature search, experiment design and execution, analysis, manuscript drafting, and review. End-to-end AI scientist systems can now produce paper-like manuscripts, but their claims are often harder to verify than their code is to run. This survey studies that gap in computational AI/ML research, where code, benchmarks, experiments, and write-ups are most visible. We screen 125 candidate works and include 35, with full-text coding of 26 entries: 24 runnable systems and two study or position papers. We code seven audit dimensions: lifecycle stage, autonomy level, evaluation method, released artifacts, human-in-the-loop points, novelty verification, and result-selection disclosure. The main pattern is that code release is now common, but reproducibility-grade and claim-verification artifacts remain much less common. In the 24 runnable systems, 83 percent release code, while 38 percent release seeds or execution traces and 38 percent report any novelty-verification method. Among nine closed-loop L4 systems, seven are mechanical reruns and one is author-claimed without an external check; no LLM-era system in the corpus demonstrates an externally validated in-loop oracle under our coding rule. We contribute a coded corpus, a lifecycle-by-autonomy map, an auditability-gap analysis, and a reviewer-facing reporting checklist. The survey argues that the field's central bottleneck is no longer only whether agents can complete research tasks, but whether reviewers can verify the claims those agents produce.

Jigsaw puzzle solving has been an increasingly popular task in the computer vision research community. Recent works have utilized cutting-edge architectures and computational approaches to reassemble groups of pieces into a coherent image, while achieving increasingly good results on well established datasets. However, most of these approaches share a common, restricting setting: operating solely on strictly square puzzle pieces. In this work, we introduce GAP, a set of novel jigsaw puzzles datasets containing synthetic, heavily eroded pieces of unrestricted shapes, generated by a learned distribution of real-world archaeological fragments. We also introduce PuzzleFlow, a novel ViT and Flow-Matching based framework for jigsaw puzzle solving, capable of handling complex puzzle pieces and demonstrating superior performance on GAP when compared to both classic and recent prominent works in this domain.

Diffusion models have demonstrated remarkable capabilities in generating high-quality samples and enhancing performance across diverse domains through Classifier-Free Guidance (CFG). However, the quality of generated samples is highly sensitive to the selection of the guidance weight. In this work, we identify a critical ``training-inference gap'' and we argue that it is the presence of this gap that undermines the performance of conditional generation and renders outputs highly sensitive to the guidance weight. We quantify this gap by measuring the accumulated error during the inference stage and establish a correlation between the selection of guidance weight and minimizing this gap. Furthermore, to mitigate this gap, we propose DiffIER, an optimization-based method for high-quality generation. We demonstrate that the accumulated error can be effectively reduced by an iterative error minimization at each step during inference. By introducing this novel plug-and-play optimization framework, we enable the optimization of errors at every single inference step and enhance generation quality. Empirical results demonstrate that our proposed method outperforms baseline approaches in conditional generation tasks. Furthermore, the method achieves consistent success in text-to-image generation, image super-resolution, and text-to-speech generation, underscoring its versatility and potential for broad applications in future research.

In this work, we empirically show that updating pretrained LMs (350M, 1.3B, 2.7B) with just a few steps of Gradient Ascent Post-training (GAP) on random, unlabeled text corpora enhances its zero-shot generalization capabilities across diverse NLP tasks. Specifically, we show that GAP can allow LMs to become comparable to 2-3x times larger LMs across 12 different NLP tasks. We also show that applying GAP on out-of-distribution corpora leads to the most reliable performance improvements. Our findings indicate that GAP can be a promising method for improving the generalization capability of LMs without any task-specific fine-tuning.

How neuronal circuits achieve credit assignment remains a central unsolved question in systems neuroscience. Various studies have suggested plausible solutions for back-propagating error signals through multi-layer networks. These purely functionally motivated models assume distinct neuronal compartments to represent local error signals that determine the sign of synaptic plasticity. However, this explicit error modulation is inconsistent with phenomenological plasticity models in which the sign depends primarily on postsynaptic activity. Here we show how a plausible microcircuit model and Hebbian learning rule derived within an adaptive control theory framework can resolve this discrepancy. Assuming errors are encoded in top-down dis-inhibitory synaptic afferents, we show that error-modulated learning emerges naturally at the circuit level when recurrent inhibition explicitly influences Hebbian plasticity. The same learning rule accounts for experimentally observed plasticity in the absence of inhibition and performs comparably to back-propagation of error (BP) on several non-linearly separable benchmarks. Our findings bridge the gap between functional and experimentally observed plasticity rules and make concrete predictions on inhibitory modulation of excitatory plasticity.

Large language models (LLMs) achieve strong performance across benchmarks--from knowledge quizzes and math reasoning to web-agent tasks--but these tests occur in static settings, lacking real dynamics and uncertainty. Consequently, they evaluate isolated reasoning or problem-solving rather than decision-making under uncertainty. To address this, we introduce LiveTradeBench, a live trading environment for evaluating LLM agents in realistic and evolving markets. LiveTradeBench follows three design principles: (i) Live data streaming of market prices and news, eliminating dependence on offline backtesting and preventing information leakage while capturing real-time uncertainty; (ii) a portfolio-management abstraction that extends control from single-asset actions to multi-asset allocation, integrating risk management and cross-asset reasoning; and (iii) multi-market evaluation across structurally distinct environments--U.S. stocks and Polymarket prediction markets--differing in volatility, liquidity, and information flow. At each step, an agent observes prices, news, and its portfolio, then outputs percentage allocations that balance risk and return. Using LiveTradeBench, we run 50-day live evaluations of 21 LLMs across families. Results show that (1) high LMArena scores do not imply superior trading outcomes; (2) models display distinct portfolio styles reflecting risk appetite and reasoning dynamics; and (3) some LLMs effectively leverage live signals to adapt decisions. These findings expose a gap between static evaluation and real-world competence, motivating benchmarks that test sequential decision making and consistency under live uncertainty.

Predicting real-world events from live market signals demands systems that fuse qualitative news with quantitative order-book dynamics under strict temporal discipline -- a challenge existing benchmarks fail to capture. We present PolyBench, a multimodal benchmark derived from Polymarket that records point-in-time cross-sections of 38,666 binary prediction markets spanning 4,997 events, synchronously coupling each snapshot with a Central Limit Order Book (CLOB) state and a real-time news stream. Using PolyBench, we evaluate seven state-of-the-art Large Language Models -- spanning open- and closed-source families -- generating 36,165 predictions under identical, timestamp-locked market states collected between February 6 and 12, 2026. Our multidimensional framework assesses directional accuracy, our proposed Confidence-Weighted Return (CWR), Annualized Percentage Yield (APY), and Sharpe ratio via realistic order-book execution simulation. The results reveal a pronounced performance divergence: only two of seven models achieve positive financial returns -- MiMo-V2-Flash at 17.6% CWR and Gemini-3-Flash at 6.2% CWR -- while the remaining five incur losses despite uniformly high stated confidence. These findings highlight the gap between surface-level language fluency and genuine probabilistic reasoning under live market uncertainty, and establish PolyBench as a contamination-proof, financially-grounded evaluation standard for future LLM research. Our dataset and code available at \href{https://github.com/PolyBench/PolyBench{https://github.com/PolyBench/PolyBench}}.

Can we trust the reasoning traces that large reasoning models (LRMs) produce? We investigate whether these traces faithfully reflect what drives model outputs, and whether models will honestly report their influence. We introduce Thought Injection, a method that injects synthetic reasoning snippets into a model's <think> trace, then measures whether the model follows the injected reasoning and acknowledges doing so. Across 45,000 samples from three LRMs, we find that injected hints reliably alter outputs, confirming that reasoning traces causally shape model behavior. However, when asked to explain their changed answers, models overwhelmingly refuse to disclose the influence: overall non-disclosure exceeds 90% for extreme hints across 30,000 follow-up samples. Instead of acknowledging the injected reasoning, models fabricate aligned-appearing but unrelated explanations. Activation analysis reveals that sycophancy- and deception-related directions are strongly activated during these fabrications, suggesting systematic patterns rather than incidental failures. Our findings reveal a gap between the reasoning LRMs follow and the reasoning they report, raising concern that aligned-appearing explanations may not be equivalent to genuine alignment.

GPT-6 Astra exhibits a remarkable ability to generate numerical robot actions, extending its role beyond high-level planning. To assess Astra's capabilities as general-purpose embodied policies, we conduct comprehensive evaluations across six domains, examining direct control, cooperation with learned policies, and feedback-driven adaptation. In gripper manipulation, Astra can correct task targets and prepare contact conditions for subsequent policy execution; hybrid control with π0.5 achieves 48% success on the evaluated RoboDojo subset. In dexterous manipulation, hybrid control achieves 50% success in ten experience-guided DexJoCo trials, while direct in-hand control struggles to coordinate finger contacts. In mobile manipulation, hybrid control reaches 38.7% success on the evaluated RoboCasa365. In navigation, Astra leads our local comparisons, reaching 92% success on RxR instruction following and 82% on HM3D object search, although search incurs substantial detours. In locomotion, dense motion-reference generation remains unreliable: none of five sequential attempts on a single obstacle course reaches the goal, despite improvements in stability and forward progress. In humanoid loco-manipulation, Astra exceeds baseline methods on 13 of 30 HumanoidBench tasks with pretrained whole-body controllers. These findings reveal a gap between useful task decisions and reliable physical control. Inference latency further constrains practical control: across 50 RoboDojo instances per condition, policy-assisted and direct control consume 624.8 million and 1.132 billion tokens. A 30-second locomotion run requires 250 model calls averaging 39.86 seconds each, with physics paused during inference.

Perceiving and remembering the visual world is fundamental to navigating and interacting with our environment. Current 4D foundation models, such as camera-controllable video models or 4D reconstruction models, can perceive and reconstruct dynamic environments, but how well they remember what they have perceived remains an open question. Existing benchmarks largely rely on pixel-level metrics and lack ground truth for objects once they leave the field of view, making them unable to evaluate visual memory in an object-centric manner against references. To fill this gap, we introduce PersistBench, a dataset and metric suite that leverages 360° videos as omniscient ground truth and proposes three evaluation aspects: object permanence, motion continuity, and appearance preservation. Evaluating various models across diverse categories reveals that current models can only maintain short-term consistency that degrades significantly once objects leave the field of view. Our findings highlight the gap between current model capabilities and robust visual memory ("seeing is not remembering"), providing guidance for future development of 4D foundation models. Dataset and code are available on the project page: https://guangzhaohe.com/persistbench.

LIBERO has emerged as a widely adopted benchmark for evaluating robotic manipulation policies. However, strong performance on LIBERO does not necessarily indicate robust embodied capabilities, as models may exploit benchmark-specific regularities rather than generalize to diverse operating conditions. In this work, we introduce LIBERO-Pro2.0, a comprehensive benchmark for systematically evaluating the robustness of embodied foundation models under distribution shifts. LIBERO-Pro2.0 encompasses 22 static perturbations and 20 runtime interventions across six dimensions: Condition, Environment, Observation, Execution, Robot State Configuration, and Language, yielding 1,648 validated shifted task instances. To support robustness-oriented learning, we further collect 39,390 training trajectories through model-generated and human demonstrations. All shifted task instances undergo six rounds of iterative human review and correction to ensure physical plausibility and task validity. We evaluate 17 representative robot policies spanning vision-language-action models, world-action models, and robustness-oriented policies. Our results reveal substantial robustness gaps across all evaluated paradigms: under runtime distribution shifts, the evaluated models exhibit an average performance drop of 26.5 percentage points, with individual drops reaching 32.5 points. These findings highlight the gap between nominal benchmark performance and robustness under distribution shifts, underscoring the need for systematic robustness evaluation of embodied foundation models.

Recent advancements in AI have led to the development of large multimodal models (LMMs) capable of processing complex tasks involving joint reasoning over text and visual content in the image (e.g., navigating maps in public places). This paper introduces ConTextual, a novel benchmark comprising instructions designed explicitly to evaluate LMMs' ability to perform context-sensitive text-rich visual reasoning. ConTextual emphasizes diverse real-world scenarios (e.g., time-reading, navigation, shopping and more) demanding a deeper understanding of the interactions between textual and visual elements. Our findings reveal a significant performance gap of 30.8% between the best-performing LMM, GPT-4V(ision), and human capabilities using human evaluation indicating substantial room for improvement in context-sensitive text-rich visual reasoning. Notably, while GPT-4V excelled in abstract categories like meme and quote interpretation, its overall performance still lagged behind humans. In addition to human evaluations, we also employed automatic evaluation metrics using GPT-4, uncovering similar trends in performance disparities. We also perform a fine-grained evaluation across diverse visual contexts and provide qualitative analysis which provides a robust framework for future advancements in the LMM design. https://con-textual.github.io/

Recent advances in Large Language Models (LLMs) have highlighted the need for robust, comprehensive, and challenging benchmarks. Yet, research on evaluating their Emotional Intelligence (EI) is considerably limited. Existing benchmarks have two major shortcomings: first, they mainly focus on emotion recognition, neglecting essential EI capabilities such as emotion regulation and thought facilitation through emotion understanding; second, they are primarily constructed from existing datasets, which include frequent patterns, explicit information, and annotation errors, leading to unreliable evaluation. We propose EmoBench, a benchmark that draws upon established psychological theories and proposes a comprehensive definition for machine EI, including Emotional Understanding and Emotional Application. EmoBench includes a set of 400 hand-crafted questions in English and Chinese, which are meticulously designed to require thorough reasoning and understanding. Our findings reveal a considerable gap between the EI of existing LLMs and the average human, highlighting a promising direction for future research. Our code and data will be publicly available from https://github.com/Sahandfer/EmoBench.

We study the depth of grade-school math (GSM) problem-solving capabilities of LLMs. To this end, we evaluate their performance on pairs of existing math word problems together so that the answer to the second problem depends on correctly answering the first problem. Our findings reveal a significant reasoning gap in most LLMs, that is performance difference between solving the compositional pairs and solving each question independently. This gap is more pronounced in smaller, more cost-efficient, and math-specialized models. Moreover, instruction-tuning recipes and code generation have varying effects across LLM sizes, while finetuning on GSM can lead to task overfitting. Our analysis indicates that large reasoning gaps are not because of test-set leakage, but due to distraction from additional context and poor second-hop reasoning. Overall, LLMs exhibit systematic differences in their reasoning abilities, despite what their performance on standard benchmarks indicates.

Current benchmarks for evaluating Vision Language Models (VLMs) often fall short in thoroughly assessing model abilities to understand and process complex visual and textual content. They typically focus on simple tasks that do not require deep reasoning or the integration of multiple data modalities to solve an original problem. To address this gap, we introduce the PARROT-360V Benchmark, a novel and comprehensive benchmark featuring 2487 challenging visual puzzles designed to test VLMs on complex visual reasoning tasks. We evaluated leading models: GPT-4o, Claude-3.5-Sonnet, and Gemini-1.5-Pro, using PARROT-360V to assess their capabilities in combining visual clues with language skills to solve tasks in a manner akin to human problem-solving. Our findings reveal a notable performance gap: state-of-the-art models scored between 28 to 56 percentage on our benchmark, significantly lower than their performance on popular benchmarks. This underscores the limitations of current VLMs in handling complex, multi-step reasoning tasks and highlights the need for more robust evaluation frameworks to advance the field.

Large Language Models (LLMs) have shown impressive performance on various benchmarks, yet their ability to engage in deliberate reasoning remains questionable. We present NYT-Connections, a collection of 358 simple word classification puzzles derived from the New York Times Connections game. This benchmark is designed to penalize quick, intuitive "System 1" thinking, isolating fundamental reasoning skills. We evaluated six recent LLMs, a simple machine learning heuristic, and humans across three configurations: single-attempt, multiple attempts without hints, and multiple attempts with contextual hints. Our findings reveal a significant performance gap: even top-performing LLMs like GPT-4 fall short of human performance by nearly 30%. Notably, advanced prompting techniques such as Chain-of-Thought and Self-Consistency show diminishing returns as task difficulty increases. NYT-Connections uniquely combines linguistic isolation, resistance to intuitive shortcuts, and regular updates to mitigate data leakage, offering a novel tool for assessing LLM reasoning capabilities.

Recent advances in large language models (LLMs) have introduced the novel paradigm of using LLMs as judges, where an LLM evaluates and scores the outputs of another LLM, which often correlates highly with human preferences. However, the use of LLM-as-a-judge has been primarily studied in English. In this paper, we evaluate this framework in Russian by introducing the Russian Error tyPes Annotation dataset (REPA), a dataset of 1k user queries and 2k LLM-generated responses. Human annotators labeled each response pair expressing their preferences across ten specific error types, as well as selecting an overall preference. We rank six generative LLMs across the error types using three rating systems based on human preferences. We also evaluate responses using eight LLM judges in zero-shot and few-shot settings. We describe the results of analyzing the judges and position and length biases. Our findings reveal a notable gap between LLM judge performance in Russian and English. However, rankings based on human and LLM preferences show partial alignment, suggesting that while current LLM judges struggle with fine-grained evaluation in Russian, there is potential for improvement.

Multimodal Large Language Models (MLLMs) have demonstrated impressive 2D image/video understanding capabilities. However, there are no publicly standardized benchmarks to assess the abilities of MLLMs in understanding the 4D objects (3D objects with temporal evolution over time). In this paper, we introduce 4D-Bench, the first benchmark to evaluate the capabilities of MLLMs in 4D object understanding, featuring tasks in 4D object Question Answering (4D object QA) and 4D object captioning. 4D-Bench provides 4D objects with diverse categories, high-quality annotations, and tasks necessitating multi-view spatial-temporal understanding, different from existing 2D image/video-based benchmarks. With 4D-Bench, we evaluate a wide range of open-source and closed-source MLLMs. The results from the 4D object captioning experiment indicate that MLLMs generally exhibit weaker temporal understanding compared to their appearance understanding, notably, while open-source models approach closed-source performance in appearance understanding, they show larger performance gaps in temporal understanding. 4D object QA yields surprising findings: even with simple single-object videos, MLLMs perform poorly, with state-of-the-art GPT-4o achieving only 63\% accuracy compared to the human baseline of 91\%. These findings highlight a substantial gap in 4D object understanding and the need for further advancements in MLLMs.

In the context of the rising interest in code language models (code LMs) and vulnerability detection, we study the effectiveness of code LMs for detecting vulnerabilities. Our analysis reveals significant shortcomings in existing vulnerability datasets, including poor data quality, low label accuracy, and high duplication rates, leading to unreliable model performance in realistic vulnerability detection scenarios. Additionally, the evaluation methods used with these datasets are not representative of real-world vulnerability detection. To address these challenges, we introduce PrimeVul, a new dataset for training and evaluating code LMs for vulnerability detection. PrimeVul incorporates a novel set of data labeling techniques that achieve comparable label accuracy to human-verified benchmarks while significantly expanding the dataset. It also implements a rigorous data de-duplication and chronological data splitting strategy to mitigate data leakage issues, alongside introducing more realistic evaluation metrics and settings. This comprehensive approach aims to provide a more accurate assessment of code LMs' performance in real-world conditions. Evaluating code LMs on PrimeVul reveals that existing benchmarks significantly overestimate the performance of these models. For instance, a state-of-the-art 7B model scored 68.26% F1 on BigVul but only 3.09% F1 on PrimeVul. Attempts to improve performance through advanced training techniques and larger models like GPT-3.5 and GPT-4 were unsuccessful, with results akin to random guessing in the most stringent settings. These findings underscore the considerable gap between current capabilities and the practical requirements for deploying code LMs in security roles, highlighting the need for more innovative research in this domain.

Cancer patients are increasingly turning to large language models (LLMs) as a new form of internet search for medical information, making it critical to assess how well these models handle complex, personalized questions. However, current medical benchmarks focus on medical exams or consumer-searched questions and do not evaluate LLMs on real patient questions with detailed clinical contexts. In this paper, we first evaluate LLMs on cancer-related questions drawn from real patients, reviewed by three hematology oncology physicians. While responses are generally accurate, with GPT-4-Turbo scoring 4.13 out of 5, the models frequently fail to recognize or address false presuppositions in the questions-posing risks to safe medical decision-making. To study this limitation systematically, we introduce Cancer-Myth, an expert-verified adversarial dataset of 585 cancer-related questions with false presuppositions. On this benchmark, no frontier LLM -- including GPT-4o, Gemini-1.Pro, and Claude-3.5-Sonnet -- corrects these false presuppositions more than 30% of the time. Even advanced medical agentic methods do not prevent LLMs from ignoring false presuppositions. These findings expose a critical gap in the clinical reliability of LLMs and underscore the need for more robust safeguards in medical AI systems.

Recent advances in large language models (LLMs) have fueled the vision of automated scientific discovery, often called AI Co-Scientists. To date, prior work casts these systems as generative co-authors responsible for crafting hypotheses, synthesizing code, or drafting manuscripts. In this work, we explore a complementary application: using LLMs as verifiers to automate the academic verification of scientific manuscripts. To that end, we introduce SPOT, a dataset of 83 published papers paired with 91 errors significant enough to prompt errata or retraction, cross-validated with actual authors and human annotators. Evaluating state-of-the-art LLMs on SPOT, we find that none surpasses 21.1\% recall or 6.1\% precision (o3 achieves the best scores, with all others near zero). Furthermore, confidence estimates are uniformly low, and across eight independent runs, models rarely rediscover the same errors, undermining their reliability. Finally, qualitative analysis with domain experts reveals that even the strongest models make mistakes resembling student-level misconceptions derived from misunderstandings. These findings highlight the substantial gap between current LLM capabilities and the requirements for dependable AI-assisted academic verification.

Current medical retrieval benchmarks primarily emphasize lexical or shallow semantic similarity, overlooking the reasoning-intensive demands that are central to clinical decision-making. In practice, physicians often retrieve authoritative medical evidence to support diagnostic hypotheses. Such evidence typically aligns with an inferred diagnosis rather than the surface form of a patient's symptoms, leading to low lexical or semantic overlap between queries and relevant documents. To address this gap, we introduce R2MED, the first benchmark explicitly designed for reasoning-driven medical retrieval. It comprises 876 queries spanning three tasks: Q&A reference retrieval, clinical evidence retrieval, and clinical case retrieval. These tasks are drawn from five representative medical scenarios and twelve body systems, capturing the complexity and diversity of real-world medical information needs. We evaluate 15 widely-used retrieval systems on R2MED and find that even the best model achieves only 31.4 nDCG@10, demonstrating the benchmark's difficulty. Classical re-ranking and generation-augmented retrieval methods offer only modest improvements. Although large reasoning models improve performance via intermediate inference generation, the best results still peak at 41.4 nDCG@10. These findings underscore a substantial gap between current retrieval techniques and the reasoning demands of real clinical tasks. We release R2MED as a challenging benchmark to foster the development of next-generation medical retrieval systems with enhanced reasoning capabilities. Data and code are available at https://github.com/R2MED/R2MED

While AI agents hold transformative potential in business, effective performance benchmarking is hindered by the scarcity of public, realistic business data on widely used platforms. Existing benchmarks often lack fidelity in their environments, data, and agent-user interactions, with limited coverage of diverse business scenarios and industries. To address these gaps, we introduce CRMArena-Pro, a novel benchmark for holistic, realistic assessment of LLM agents in diverse professional settings. CRMArena-Pro expands on CRMArena with nineteen expert-validated tasks across sales, service, and 'configure, price, and quote' processes, for both Business-to-Business and Business-to-Customer scenarios. It distinctively incorporates multi-turn interactions guided by diverse personas and robust confidentiality awareness assessments. Experiments reveal leading LLM agents achieve only around 58% single-turn success on CRMArena-Pro, with performance dropping significantly to approximately 35% in multi-turn settings. While Workflow Execution proves more tractable for top agents (over 83% single-turn success), other evaluated business skills present greater challenges. Furthermore, agents exhibit near-zero inherent confidentiality awareness; though targeted prompting can improve this, it often compromises task performance. These findings highlight a substantial gap between current LLM capabilities and enterprise demands, underscoring the need for advancements in multi-turn reasoning, confidentiality adherence, and versatile skill acquisition.

Multimodal large language models (MLLMs) are increasingly deployed in open-ended, real-world environments where inputs are messy, underspecified, and not always trustworthy. Unlike curated benchmarks, these settings frequently involve instructions that refer to missing objects or contradictory facts, rely on ambiguous references, or request infeasible actions. In such cases, success hinges not on task execution alone, but on a model's ability to detect when something is silently wrong. This paper presents a systematic analysis of how current MLLMs handle such implicit reasoning scenarios: cases where the flaw is not explicitly stated but must be inferred from context. Using a curated diagnostic suite spanning four categories of real-world failure modes, we evaluate six MLLMs, including o3 and GPT-4o, and find that models frequently fail to surface hidden issues, even when they possess the necessary perceptual and reasoning skills. Explicit prompting reveals that the underlying capabilities exist but are often suppressed in favor of user compliance. We further show that simple inference-time interventions, such as cautious persona prompting and, in particular, requiring a clarifying question, can dramatically recover performance. Our findings highlight a persistent gap between reasoning competence and behavioral compliance in current MLLMs and suggest practical strategies for making these models more trustworthy in underconstrained environments.

The advancement of Large Language Models (LLMs) has led to significant improvements in various service domains, including search, recommendation, and chatbot applications. However, applying state-of-the-art (SOTA) research to industrial settings presents challenges, as it requires maintaining flexible conversational abilities while also strictly complying with service-specific constraints. This can be seen as two conflicting requirements due to the probabilistic nature of LLMs. In this paper, we propose our approach to addressing this challenge and detail the strategies we employed to overcome their inherent limitations in real-world applications. We conduct a practical case study of a conversational agent designed for the e-commerce domain, detailing our implementation workflow and optimizations. Our findings provide insights into bridging the gap between academic research and real-world application, introducing a framework for developing scalable, controllable, and reliable AI-driven agents.

We introduce Drivelology, a unique linguistic phenomenon characterised as "nonsense with depth", utterances that are syntactically coherent yet pragmatically paradoxical, emotionally loaded, or rhetorically subversive. While such expressions may resemble surface-level nonsense, they encode implicit meaning requiring contextual inference, moral reasoning, or emotional interpretation. We find that current large language models (LLMs), despite excelling at many natural language processing (NLP) tasks, consistently fail to grasp the layered semantics of Drivelological text. To investigate this, we construct a small but diverse benchmark dataset of over 1,200 meticulously curated examples, with select instances in English, Mandarin, Spanish, French, Japanese, and Korean. Annotation was especially challenging: each of the examples required careful expert review to verify that it truly reflected Drivelological characteristics. The process involved multiple rounds of discussion and adjudication to address disagreements, highlighting the subtle and subjective nature of the Drivelology. We evaluate a range of LLMs on classification, generation, and reasoning tasks. Our results reveal clear limitations of LLMs: models often confuse Drivelology with shallow nonsense, produce incoherent justifications, or miss the implied rhetorical function altogether. These findings highlight a deeper representational gap in LLMs' pragmatic understanding and challenge the assumption that statistical fluency implies cognitive comprehension. We release our dataset and code to facilitate further research in modelling linguistic depth beyond surface-level coherence.

AIOps (Artificial Intelligence for IT Operations) solutions leverage the tremendous amount of data produced during the operation of large-scale systems and machine learning models to assist software practitioners in their system operations. Existing AIOps solutions usually maintain AIOps models against concept drift through periodical retraining, despite leaving a pile of discarded historical models that may perform well on specific future data. Other prior works propose dynamically selecting models for prediction tasks from a set of candidate models to optimize the model performance. However, there is no prior work in the AIOps area that assesses the use of model selection mechanisms on historical models to improve model performance or robustness. To fill the gap, we evaluate several model selection mechanisms by assessing their capabilities in selecting the optimal AIOps models that were built in the past to make predictions for the target data. We performed a case study on three large-scale public operation datasets: two trace datasets from the cloud computing platforms of Google and Alibaba, and one disk stats dataset from the BackBlaze cloud storage data center. We observe that the model selection mechnisms utilizing temporal adjacency tend to have a better performance and can prevail the periodical retraining approach. Our findings also highlight a performance gap between existing model selection mechnisms and the theoretical upper bound which may motivate future researchers and practitioners in investigating more efficient and effective model selection mechanisms that fit in the context of AIOps.

We study the problem of controlling the difficulty level of text generated by Large Language Models (LLMs) for contexts where end-users are not fully proficient, such as language learners. Using a novel framework, we evaluate the effectiveness of several key approaches for this task, including few-shot prompting, supervised finetuning, and reinforcement learning (RL), utilising both GPT-4 and open source alternatives like LLama2-7B and Mistral-7B. Our findings reveal a large performance gap between GPT-4 and the open source models when using prompt-based strategies. However, we show how to bridge this gap with a careful combination of finetuning and RL alignment. Our best model, CALM (CEFR-Aligned Language Model), surpasses the performance of GPT-4 and other strategies, at only a fraction of the cost. We further validate the quality of our results through a small-scale human study.

Large Language Models (LLMs) have recently demonstrated strong capabilities in translating natural language into database queries, especially when dealing with complex graph-structured data. However, real-world queries often contain inherent ambiguities, and the interconnected nature of graph structures can amplify these challenges, leading to unintended or incorrect query results. To systematically evaluate LLMs on this front, we propose a taxonomy of graph-query ambiguities, comprising three primary types: Attribute Ambiguity, Relationship Ambiguity, and Attribute-Relationship Ambiguity, each subdivided into Same-Entity and Cross-Entity scenarios. We introduce AmbiGraph-Eval, a novel benchmark of real-world ambiguous queries paired with expert-verified graph query answers. Evaluating 9 representative LLMs shows that even top models struggle with ambiguous graph queries. Our findings reveal a critical gap in ambiguity handling and motivate future work on specialized resolution techniques.

Large language models (LLMs) have advanced the automation of data science workflows. Yet it remains unclear whether they can critically leverage external domain knowledge as human data scientists do in practice. To answer this question, we introduce AssistedDS (Assisted Data Science), a benchmark designed to systematically evaluate how LLMs handle domain knowledge in tabular prediction tasks. AssistedDS features both synthetic datasets with explicitly known generative mechanisms and real-world Kaggle competitions, each accompanied by curated bundles of helpful and adversarial documents. These documents provide domain-specific insights into data cleaning, feature engineering, and model selection. We assess state-of-the-art LLMs on their ability to discern and apply beneficial versus harmful domain knowledge, evaluating submission validity, information recall, and predictive performance. Our results demonstrate three key findings: (1) LLMs frequently exhibit an uncritical adoption of provided information, significantly impairing their predictive performance when adversarial content is introduced, (2) helpful guidance is often insufficient to counteract the negative influence of adversarial information, and (3) in Kaggle datasets, LLMs often make errors in handling time-series data, applying consistent feature engineering across different folds, and interpreting categorical variables correctly. These findings highlight a substantial gap in current models' ability to critically evaluate and leverage expert knowledge, underscoring an essential research direction for developing more robust, knowledge-aware automated data science systems.

Recent years have witnessed significant progress in Unified Multimodal Models, yet a fundamental question remains: Does understanding truly inform generation? To investigate this, we introduce UniSandbox, a decoupled evaluation framework paired with controlled, synthetic datasets to avoid data leakage and enable detailed analysis. Our findings reveal a significant understanding-generation gap, which is mainly reflected in two key dimensions: reasoning generation and knowledge transfer. Specifically, for reasoning generation tasks, we observe that explicit Chain-of-Thought (CoT) in the understanding module effectively bridges the gap, and further demonstrate that a self-training approach can successfully internalize this ability, enabling implicit reasoning during generation. Additionally, for knowledge transfer tasks, we find that CoT assists the generative process by helping retrieve newly learned knowledge, and also discover that query-based architectures inherently exhibit latent CoT-like properties that affect this transfer. UniSandbox provides preliminary insights for designing future unified architectures and training strategies that truly bridge the gap between understanding and generation. Code and data are available at https://github.com/PKU-YuanGroup/UniSandBox

We propose Perceptual Taxonomy, a structured process of scene understanding that first recognizes objects and their spatial configurations, then infers task-relevant properties such as material, affordance, function, and physical attributes to support goal-directed reasoning. While this form of reasoning is fundamental to human cognition, current vision-language benchmarks lack comprehensive evaluation of this ability and instead focus on surface-level recognition or image-text alignment. To address this gap, we introduce Perceptual Taxonomy, a benchmark for physically grounded visual reasoning. We annotate 3173 objects with four property families covering 84 fine-grained attributes. Using these annotations, we construct a multiple-choice question benchmark with 5802 images across both synthetic and real domains. The benchmark contains 28033 template-based questions spanning four types (object description, spatial reasoning, property matching, and taxonomy reasoning), along with 50 expert-crafted questions designed to evaluate models across the full spectrum of perceptual taxonomy reasoning. Experimental results show that leading vision-language models perform well on recognition tasks but degrade by 10 to 20 percent on property-driven questions, especially those requiring multi-step reasoning over structured attributes. These findings highlight a persistent gap in structured visual understanding and the limitations of current models that rely heavily on pattern matching. We also show that providing in-context reasoning examples from simulated scenes improves performance on real-world and expert-curated questions, demonstrating the effectiveness of perceptual-taxonomy-guided prompting.

We present ShoppingComp, a challenging real-world benchmark for rigorously evaluating LLM-powered shopping agents on three core capabilities: precise product retrieval, expert-level report generation, and safety critical decision making. Unlike prior e-commerce benchmarks, ShoppingComp introduces highly complex tasks under the principle of guaranteeing real products and ensuring easy verifiability, adding a novel evaluation dimension for identifying product safety hazards alongside recommendation accuracy and report quality. The benchmark comprises 120 tasks and 1,026 scenarios, curated by 35 experts to reflect authentic shopping needs. Results reveal stark limitations of current LLMs: even state-of-the-art models achieve low performance (e.g., 11.22% for GPT-5, 3.92% for Gemini-2.5-Flash). These findings highlight a substantial gap between research benchmarks and real-world deployment, where LLMs make critical errors such as failure to identify unsafe product usage or falling for promotional misinformation, leading to harmful recommendations. ShoppingComp fills the gap and thus establishes a new standard for advancing reliable and practical agents in e-commerce.

Modern multimodal large language models often claim "video understanding," yet most evaluations use muted videos or simply discard audio. We ask a direct question: how much does audio actually matter for contemporary Video-LLMs and the benchmarks that certify them? We audit widely used suites and observe that many items are even solvable from a single frame, rendering audio largely redundant. Building on LLaVA-OneVision architecture, we attach a speech/audio encoder (e.g., Whisper) and analyze when audio helps, while addressing audio token explosion with a lightweight Mamba-based state-space token compressor. We find that audio yields minimal gains on recent video benchmarks but is decisive on curated, audio-sensitive subsets. To enable faithful evaluation, we release AVQA-Hard and Music-AVQA-Hard, our model, and code. Our findings surface a growing gap between current academic practice and real-world expectations, and provide practical tools for scalable audio-visual Video-LLMs. We will fully open-source our work at https://github.com/naver-ai/LLaVA-AV-SSM.