LLM 관련 주요 논문 - 2025-12-01

1. On the Limits of Innate Planning in Large Language Models

Authors: Charles Schepanowski , Charles Ling
URL: https://arxiv.org/abs/2511.21591
Abstract:

Large language models (LLMs) achieve impressive results on many benchmarks, yet their capacity for planning and stateful reasoning remains unclear. We study these abilities directly, without code execution or other tools, using the 8-puzzle: a classic task that requires state tracking and goal-directed planning while allowing precise, step-by-step evaluation. Four models are tested under common prompting conditions (Zero-Shot, Chain-of-Thought, Algorithm-of-Thought) and with tiered corrective feedback. Feedback improves success rates for some model-prompt combinations, but many successful runs are long, computationally expensive, and indirect. We then examine the models with an external move validator that provides only valid moves. Despite this level of assistance, none of the models solve any puzzles in this setting. Qualitative analysis reveals two dominant deficits across all models: (1) brittle internal state representations, leading to frequent invalid moves, and (2) weak heuristic planning, with models entering loops or selecting actions that do not reduce the distance to the goal state. These findings indicate that, in the absence of external tools such as code interpreters, current LLMs have substantial limitations in planning and that further progress may require mechanisms for maintaining explicit state and performing structured search.

2. Self-Transparency Failures in Expert-Persona LLMs: A Large-Scale Behavioral Audit

Authors: Alex Diep
URL: https://arxiv.org/abs/2511.21569
Abstract:

If a language model cannot reliably disclose its AI identity in expert contexts, users cannot trust its competence boundaries. This study examines self-transparency in models assigned professional personas within high-stakes domains where false expertise risks user harm. Using a common-garden design, sixteen open-weight models (4B–671B parameters) were audited across 19,200 trials. Models exhibited sharp domain-specific inconsistency: a Financial Advisor persona elicited 30.8% disclosure initially, while a Neurosurgeon persona elicited only 3.5%. This creates preconditions for a “Reverse Gell-Mann Amnesia” effect, where transparency in some domains leads users to overgeneralize trust to contexts where disclosure fails. Disclosure ranged from 2.8% to 73.6%, with a 14B model reaching 61.4% while a 70B produced just 4.1%. Model identity predicted behavior better than parameter count ($\Delta R_{adj}^{2} = 0.359$ vs 0.018). Reasoning optimization actively suppressed self-transparency in some models, with reasoning variants showing up to 48.4% lower disclosure than base counterparts. Bayesian validation with Rogan–Gladen correction confirmed robustness to measurement error ($\kappa = 0.908$). These findings demonstrate transparency reflects training factors rather than scale. Organizations cannot assume safety properties transfer to deployment contexts, requiring deliberate behavior design and empirical verification.

3. Pessimistic Verification for Open Ended Math Questions

Authors: Yanxing Huang , Zihan Tang , Zejin Lin , Peng Li , Yang Liu
URL: https://arxiv.org/abs/2511.21522
Abstract:

The key limitation of the verification performance lies in the ability of error detection. With this intuition we designed several variants of pessimistic verification, which are simple workflows that could significantly improve the verification of open-ended math questions. In pessimistic verification we construct multiple parallel verifications for the same proof, and the proof is deemed incorrect if any one of them reports an error. This simple technique significantly improves the performance across many math verification benchmarks without incurring substantial computational resources. Its token efficiency even surpassed extended long-CoT in test-time scaling. Our case studies further indicate that the majority of false negatives in stronger models are actually caused by annotation errors in the original dataset, so our method’s performance is in fact underestimated. Self-verification for mathematical problems can effectively improve the reliability and performance of language model outputs, and it also plays a critical role in enabling long-horizon mathematical tasks. We believe that research on pessimistic verification will help enhance the mathematical capabilities of language models across a wide range of tasks.

4. SpatialBench: Benchmarking Multimodal Large Language Models for Spatial Cognition

Authors: Peiran Xu , Sudong Wang , Yao Zhu , Jianing Li , Yunjian Zhang
URL: https://arxiv.org/abs/2511.21471
Abstract:

Spatial cognition is fundamental to real-world multimodal intelligence, allowing models to effectively interact with the physical environment. While multimodal large language models (MLLMs) have made significant strides, existing benchmarks often oversimplify spatial cognition, reducing it to a single-dimensional metric, which fails to capture the hierarchical structure and interdependence of spatial abilities. To address this gap, we propose a hierarchical spatial cognition framework that decomposes spatial intelligence into five progressively complex levels from basic observation to high-level planning. Building upon this taxonomy, we construct SpatialBench, a large-scale, fine-grained benchmark covering 15 tasks aligned with these cognitive levels. To provide a unified evaluation across heterogeneous tasks, we further introduce a high-level capability-oriented metric that reliably assesses a model’s overall spatial reasoning ability. Extensive experiments over massive MLLMs reveal distinct performance stratification across cognitive levels: models exhibit strong perceptual grounding yet remain limited in symbolic reasoning, causal inference, and planning. Additional human tests demonstrate that humans perform selective, goal-directed abstraction, while MLLMs tend to over-attend to surface details without coherent spatial intent. Our work establishes the first systematic framework for measuring hierarchical spatial cognition in MLLMs, laying the foundation for future spatially intelligent systems.

5. MADRA: Multi-Agent Debate for Risk-Aware Embodied Planning

Authors: Junjian Wang , Lidan Zhao , Xi Sheryl Zhang
URL: https://arxiv.org/abs/2511.21460
Abstract:

Ensuring the safety of embodied AI agents during task planning is critical for real-world deployment, especially in household environments where dangerous instructions pose significant risks. Existing methods often suffer from either high computational costs due to preference alignment training or over-rejection when using single-agent safety prompts. To address these limitations, we propose MADRA, a training-free Multi-Agent Debate Risk Assessment framework that leverages collective reasoning to enhance safety awareness without sacrificing task performance. MADRA employs multiple LLM-based agents to debate the safety of a given instruction, guided by a critical evaluator that scores responses based on logical soundness, risk identification, evidence quality, and clarity. Through iterative deliberation and consensus voting, MADRA significantly reduces false rejections while maintaining high sensitivity to dangerous tasks. Additionally, we introduce a hierarchical cognitive collaborative planning framework that integrates safety, memory, planning, and self-evolution mechanisms to improve task success rates through continuous learning. We also contribute SafeAware-VH, a benchmark dataset for safety-aware task planning in VirtualHome, containing 800 annotated instructions. Extensive experiments on AI2-THOR and VirtualHome demonstrate that our approach achieves over 90% rejection of unsafe tasks while ensuring that safe-task rejection is low, outperforming existing methods in both safety and execution efficiency. Our work provides a scalable, model-agnostic solution for building trustworthy embodied agents.

6. Prune4Web: DOM Tree Pruning Programming for Web Agent

Authors: Jiayuan Zhang , Kaiquan Chen , Zhihao Lu , Enshen Zhou , Qian Yu , Jing Zhang
URL: https://arxiv.org/abs/2511.21398
Abstract:

Web automation employs intelligent agents to execute high-level tasks by mimicking human interactions with web interfaces. Despite the capabilities of recent Large Language Model (LLM)-based web agents, navigating complex, real-world webpages efficiently remains a significant hurdle due to the prohibitively large size of Document Object Model (DOM) structures, often ranging from 10,000 to 100,000 tokens. Existing strategies typically rely on crude DOM truncation – risking the loss of critical information – or employ inefficient heuristics and separate ranking models, failing to achieve an optimal balance between precision and scalability. To address these challenges, we introduce Prune4Web, a novel paradigm that shifts DOM processing from resource-intensive LLM reading to efficient programmatic pruning. Central to our approach is DOM Tree Pruning Programming, where an LLM generates executable Python scoring scripts to dynamically filter DOM elements based on semantic cues from decomposed sub-tasks. This mechanism eliminates the need for LLMs to ingest raw, massive DOMs, instead delegating traversal and scoring to lightweight, interpretable programs. This methodology achieves a 25x to 50x reduction in candidate elements for grounding, thereby facilitating precise action localization while mitigating attention dilution. Furthermore, we propose a specialized data annotation pipeline and a two-turn dialogue training strategy that jointly optimizes the Planner, Programmatic Filter, and Grounder within a unified framework. Extensive experiments demonstrate state-of-the-art performance. Notably, on our low-level grounding task, Prune4Web dramatically improves accuracy from 46.8% to 88.28%, underscoring its efficacy in real-world web automation.

7. OVOD-Agent: A Markov-Bandit Framework for Proactive Visual Reasoning and Self-Evolving Detection

Authors: Chujie Wang , Jianyu Lu , Zhiyuan Luo , Xi Chen , Chu He
URL: https://arxiv.org/abs/2511.21064
Abstract:

Open-Vocabulary Object Detection (OVOD) aims to enable detectors to generalize across categories by leveraging semantic information. Although existing methods are pretrained on large vision-language datasets, their inference is still limited to fixed category names, creating a gap between multimodal training and unimodal inference. Previous work has shown that improving textual representation can significantly enhance OVOD performance, indicating that the textual space is still underexplored. To this end, we propose OVOD-Agent, which transforms passive category matching into proactive visual reasoning and self-evolving detection. Inspired by the Chain-of-Thought (CoT) paradigm, OVOD-Agent extends the textual optimization process into an interpretable Visual-CoT with explicit actions. OVOD’s lightweight nature makes LLM-based management unsuitable; instead, we model visual context transitions as a Weakly Markovian Decision Process (w-MDP) over eight state spaces, which naturally represents the agent’s state, memory, and interaction dynamics. A Bandit module generates exploration signals under limited supervision, helping the agent focus on uncertain regions and adapt its detection policy. We further integrate Markov transition matrices with Bandit trajectories for self-supervised Reward Model (RM) optimization, forming a closed loop from Bandit exploration to RM learning. Experiments on COCO and LVIS show that OVOD-Agent provides consistent improvements across OVOD backbones, particularly on rare categories, confirming the effectiveness of the proposed framework.

8. Towards Trustworthy Legal AI through LLM Agents and Formal Reasoning

Authors: Linze Chen , Yufan Cai , Zhe Hou , Jinsong Dong
URL: https://arxiv.org/abs/2511.21033
Abstract:

The rationality of law manifests in two forms: substantive rationality, which concerns the fairness or moral desirability of outcomes, and formal rationality, which requires legal decisions to follow explicitly stated, general, and logically coherent rules. Existing LLM-based systems excel at surface-level text analysis but lack the guarantees required for principled jurisprudence. We introduce L4M, a novel framework that combines adversarial LLM agents with SMT-solver-backed proofs to unite the interpretive flexibility of natural language with the rigor of symbolic verification. The pipeline consists of three phases: (1) Statute Formalization, where domain-specific prompts convert legal provisions into logical formulae; (2) Dual Fact and Statute Extraction, in which prosecutor- and defense-aligned LLMs independently map case narratives to fact tuples and statutes, ensuring role isolation; and (3) Solver-Centric Adjudication, where an autoformalizer compiles both parties’ arguments into logic constraints, and unsat cores trigger iterative self-critique until a satisfiable formula is achieved, which is then verbalized by a Judge-LLM into a transparent verdict and optimized sentence. Experimental results on public benchmarks show that our system surpasses advanced LLMs including GPT-o4-mini, DeepSeek-V3, and Claude 4 as well as state-of-the-art Legal AI baselines, while providing rigorous and explainable symbolic justifications.

9. ICPO: Intrinsic Confidence-Driven Group Relative Preference Optimization for Efficient Reinforcement Learning

Authors: Jinpeng Wang , Chao Li , Ting Ye , Mengyuan Zhang , Wei Liu , Jian Luan
URL: https://arxiv.org/abs/2511.21005
Abstract:

Reinforcement Learning with Verifiable Rewards (RLVR) demonstrates significant potential in enhancing the reasoning capabilities of Large Language Models (LLMs). However, existing RLVR methods are often constrained by issues such as coarse-grained rewards, reward noise, and inefficient exploration, which lead to unstable training and entropy collapse. To address this challenge, we propose the Intrinsic Confidence-Driven Group Relative Preference Optimization method (ICPO). The intuition behind it lies in the fact that the probabilities of an LLM generating different responses can inherently and directly reflect its self-assessment of the reasoning process. Inspired by the idea of preference modeling, ICPO calculates a preference advantage score for each response by comparing the relative generation probabilities of multiple responses under the same input prompt, and integrates this score with verifiable rewards to guide the exploration process. We have discovered that the preference advantage score not only alleviates the issues of coarse-grained rewards and reward noise but also effectively curbs overconfident errors, enhances the relative superiority of undervalued high-quality responses, and prevents the model from overfitting to specific strategies, thereby facilitating more thorough exploration. Comprehensive experiments across four general-domain benchmarks and three mathematical benchmarks demonstrate that ICPO steadily boosts reasoning compared to GRPO.

10. Improving Procedural Skill Explanations via Constrained Generation: A Symbolic-LLM Hybrid Architecture

Authors: Rahul Dass , Thomas Bowlin , Zebing Li , Xiao Jin , Ashok Goel
URL: https://arxiv.org/abs/2511.20942
Abstract:

In procedural skill learning, instructional explanations must convey not just steps, but the causal, goal-directed, and compositional logic behind them. Large language models (LLMs) often produce fluent yet shallow responses that miss this structure. We present Ivy, an AI coaching system that delivers structured, multi-step explanations by combining symbolic Task-Method-Knowledge (TMK) models with a generative interpretation layer-an LLM that constructs explanations while being constrained by TMK structure. TMK encodes causal transitions, goal hierarchies, and problem decompositions, and guides the LLM within explicit structural bounds. We evaluate Ivy against responses against GPT and retrieval-augmented GPT baselines using expert and independent annotations across three inferential dimensions. Results show that symbolic constraints consistently improve the structural quality of explanations for “how” and “why” questions. This study demonstrates a scalable AI for education approach that strengthens the pedagogical value of AI-generated explanations in intelligent coaching systems.

11. ENACT: Evaluating Embodied Cognition with World Modeling of Egocentric Interaction

Authors: Qineng Wang , Wenlong Huang , Yu Zhou , Hang Yin , Tianwei Bao , Jianwen Lyu , Weiyu Liu , Ruohan Zhang , Jiajun Wu , Li Fei-Fei , Manling Li
URL: https://arxiv.org/abs/2511.20937
Abstract:

Embodied cognition argues that intelligence arises from sensorimotor interaction rather than passive observation. It raises an intriguing question: do modern vision-language models (VLMs), trained largely in a disembodied manner, exhibit signs of embodied cognition? We introduce ENACT, a benchmark that casts evaluation of embodied cognition as world modeling from egocentric interaction in a visual question answering (VQA) format. Framed as a partially observable Markov decision process (POMDP) whose actions are scene graph changes, ENACT comprises two complementary sequence reordering tasks: forward world modeling (reorder shuffled observations given actions) and inverse world modeling (reorder shuffled actions given observations). While conceptually simple, solving these tasks implicitly demands capabilities central to embodied cognition-affordance recognition, action-effect reasoning, embodied awareness, and interactive, long-horizon memory from partially observable egocentric input, while avoiding low-level image synthesis that could confound the evaluation. We provide a scalable pipeline that synthesizes QA pairs from robotics simulation (BEHAVIOR) and evaluates models on 8,972 QA pairs spanning long-horizon home-scale activities. Experiments reveal a performance gap between frontier VLMs and humans that widens with interaction horizon. Models consistently perform better on the inverse task than the forward one and exhibit anthropocentric biases, including a preference for right-handed actions and degradation when camera intrinsics or viewpoints deviate from human vision. Website at this https URL .

12. Representation Interventions Enable Lifelong Unstructured Knowledge Control

Authors: Xuyuan Liu , Zhengzhang Chen , Xinshuai Dong , Yanchi Liu , Xujiang Zhao , Shengyu Chen , Haoyu Wang , Yujun Yan , Haifeng Chen
URL: https://arxiv.org/abs/2511.20892
Abstract:

Large language models (LLMs) often produce incorrect or outdated content. Updating their knowledge efficiently and accurately without costly retraining is a major challenge. This problem is especially hard for complex, unstructured knowledge in a lifelong setting, where many edits must coexist without interference. We introduce RILKE (Representation Intervention for Lifelong KnowledgE Control), a robust and scalable method that treats knowledge control as interventions within the model’s representation space. Leveraging representation-space expressiveness, we identify two properties enabling RILKE to deliver fine-grained control over complex, unstructured knowledge while maintaining general utility with frozen base weights. During training, RILKE learns paraphrase-robust and edit-localized modules that limit each update to a low-dimensional subspace to minimize cross-edit interference. In inference, a query-adaptive router selects the appropriate module to guide the model’s generation. In evaluation on knowledge editing benchmarks with LLaMA and Qwen models, RILKE is scalable to large-scale datasets, demonstrating high edit success, strong paraphrase generalization, and preserving general utility with modest memory overhead. These results show RILKE is an effective and scalable solution for lifelong knowledge control in LLMs.

13. Learning Multi-Access Point Coordination in Agentic AI Wi-Fi with Large Language Models

Authors: Yifan Fan , Le Liang , Peng Liu , Xiao Li , Ziyang Guo , Qiao Lan , Shi Jin , Wen Tong
URL: https://arxiv.org/abs/2511.20719
Abstract:

Multi-access point coordination (MAPC) is a key technology for enhancing throughput in next-generation Wi-Fi within dense overlapping basic service sets. However, existing MAPC protocols rely on static, protocol-defined rules, which limits their ability to adapt to dynamic network conditions such as varying interference levels and topologies. To address this limitation, we propose a novel Agentic AI Wi-Fi framework where each access point, modeled as an autonomous large language model agent, collaboratively reasons about the network state and negotiates adaptive coordination strategies in real time. This dynamic collaboration is achieved through a cognitive workflow that enables the agents to engage in natural language dialogue, leveraging integrated memory, reflection, and tool use to ground their decisions in past experience and environmental feedback. Comprehensive simulation results demonstrate that our agentic framework successfully learns to adapt to diverse and dynamic network environments, significantly outperforming the state-of-the-art spatial reuse baseline and validating its potential as a robust and intelligent solution for future wireless networks.

14. Cross Domain Evaluation of Multimodal Chain-of-Thought Reasoning of different datasets into the Amazon CoT Framework

Authors: Nitya Tiwari , Parv Maheshwari , Vidisha Agarwal
URL: https://arxiv.org/abs/2511.20701
Abstract:

While recent work has extended CoT to multimodal settings, achieving state-of-the-art results on science question answering benchmarks like ScienceQA, the generalizability of these approaches across diverse domains remains underexplored. This work presents a comprehensive analysis of Multimodal Chain-of-Thought (Multimodal-CoT) reasoning, evaluating its effectiveness on the A-OKVQA, OKVQA and ChartQA datasets, which requires broad commonsense and world knowledge beyond scientific reasoning. We implement the two-stage framework proposed by Zhang et al. [3], which separates rationale generation from answer inference and integrates vision features through a gated fusion mechanism with T5-based language models. Through systematic ablation studies, we analyze the contributions of vision features, rationale quality, and architectural choices. Our findings reveal that while vision integration significantly reduces hallucination in rationale generation, the effectiveness of CoT reasoning varies substantially across question types, with commonsense reasoning presenting particular challenges. This work provides practical insights for researchers implementing multimodal reasoning systems and identifies key areas for future improvement in cross-domain generalization.

15. Reasoning With a Star: A Heliophysics Dataset and Benchmark for Agentic Scientific Reasoning

Authors: Kevin Lee , Russell Spiewak , James Walsh
URL: https://arxiv.org/abs/2511.20694
Abstract:

Scientific reasoning through Large Language Models in heliophysics involves more than just recalling facts: it requires incorporating physical assumptions, maintaining consistent units, and providing clear scientific formats through coordinated approaches. To address these challenges, we present Reasoning With a Star, a newly contributed heliophysics dataset applicable to reasoning; we also provide an initial benchmarking approach. Our data are constructed from National Aeronautics and Space Administration & University Corporation for Atmospheric Research Living With a Star summer school problem sets and compiled into a readily consumable question-and-answer structure with question contexts, reasoning steps, expected answer type, ground-truth targets, format hints, and metadata. A programmatic grader checks the predictions using unit-aware numerical tolerance, symbolic equivalence, and schema validation. We benchmark a single-shot baseline and four multi-agent patterns, finding that decomposing workflows through systems engineering principles outperforms direct prompting on problems requiring deductive reasoning rather than pure inductive recall.

16. $A^2Flow:$ Automating Agentic Workflow Generation via Self-Adaptive Abstraction Operators

Authors: Mingming Zhao , Xiaokang Wei , Yuanqi Shao , Kaiwen Zhou , Lin Yang , Siwei Rao , Junhui Zhan , Zhitang Chen
URL: https://arxiv.org/abs/2511.20693
Abstract:

Large language models (LLMs) have shown strong potential in automating the design of agentic workflows. However, existing methods still rely heavily on manually predefined operators, limiting generalization and scalability. To address this issue, we propose $A^2Flow$, a fully automated framework for agentic workflow generation based on self-adaptive abstraction operators. $A^2Flow$ employs a three-stage operator extraction process: 1) Case-based Initial Operator Generation: leveraging expert demonstrations and LLM reasoning to generate case-specific operators; 2) Operator Clustering and Preliminary Abstraction: grouping similar operators across tasks to form preliminary abstractions; and 3) Deep Extraction for Abstract Execution Operators: applying long chain-of-thought prompting and multi-path reasoning to derive compact and generalizable execution operators. These operators serve as reusable building blocks for workflow construction without manual predefinition. Furthermore, we enhance node-level workflow search with an operator memory mechanism, which retains historical outputs to enrich context and improve decision-making. Experiments on general and embodied benchmarks show that $A^2Flow$ achieves a 2.4\% and 19.3\% average performance improvement and reduces resource usage by 37\% over state-of-the-art baselines. Homepage: this https URL

17. Minimizing Hyperbolic Embedding Distortion with LLM-Guided Hierarchy Restructuring

Authors: Melika Ayoughi , Pascal Mettes , Paul Groth
URL: https://arxiv.org/abs/2511.20679
Abstract:

Hyperbolic geometry is an effective geometry for embedding hierarchical data structures. Hyperbolic learning has therefore become increasingly prominent in machine learning applications where data is hierarchically organized or governed by hierarchical semantics, ranging from recommendation systems to computer vision. The quality of hyperbolic embeddings is tightly coupled to the structure of the input hierarchy, which is often derived from knowledge graphs or ontologies. Recent work has uncovered that for an optimal hyperbolic embedding, a high branching factor and single inheritance are key, while embedding algorithms are robust to imbalance and hierarchy size. To assist knowledge engineers in reorganizing hierarchical knowledge, this paper investigates whether Large Language Models (LLMs) have the ability to automatically restructure hierarchies to meet these criteria. We propose a prompt-based approach to transform existing hierarchies using LLMs, guided by known desiderata for hyperbolic embeddings. Experiments on 16 diverse hierarchies show that LLM-restructured hierarchies consistently yield higher-quality hyperbolic embeddings across several standard embedding quality metrics. Moreover, we show how LLM-guided hierarchy restructuring enables explainable reorganizations, providing justifications to knowledge engineers.

18. Revisiting Generalization Across Difficulty Levels: It’s Not So Easy

Authors: Yeganeh Kordi , Nihal V. Nayak , Max Zuo , Ilana Nguyen , Stephen H. Bach
URL: https://arxiv.org/abs/2511.21692
Abstract:

We investigate how well large language models (LLMs) generalize across different task difficulties, a key question for effective data curation and evaluation. Existing research is mixed regarding whether training on easier or harder data leads to better results, and whether those gains come on easier or harder test data. We address this question by conducting a systematic evaluation of LLMs’ generalization across models, datasets, and fine-grained groups of example difficulty. We rank examples in six datasets using the outputs of thousands of different LLMs and Item Response Theory (IRT), a well-established difficulty metric in educational testing. Unlike prior work, our difficulty ratings are therefore determined solely by the abilities of many different LLMs, excluding human opinions of difficulty. With a more objective, larger-scale, and finer-grained analysis, we show that cross-difficulty generalization is often limited; training on either easy or hard data cannot achieve consistent improvements across the full range of difficulties. These results show the importance of having a range of difficulties in both training and evaluation data for LLMs, and that taking shortcuts with respect to difficulty is risky.

19. ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration

Authors: Hongjin Su , Shizhe Diao , Ximing Lu , Mingjie Liu , Jiacheng Xu , Xin Dong , Yonggan Fu , Peter Belcak , Hanrong Ye , Hongxu Yin , Yi Dong , Evelina Bakhturina , Tao Yu , Yejin Choi , Jan Kautz , Pavlo Molchanov
URL: https://arxiv.org/abs/2511.21689
Abstract:

Large language models are powerful generalists, yet solving deep and complex problems such as those of the Humanity’s Last Exam (HLE) remains both conceptually challenging and computationally expensive. We show that small orchestrators managing other models and a variety of tools can both push the upper bound of intelligence and improve efficiency in solving difficult agentic tasks. We introduce ToolOrchestra, a method for training small orchestrators that coordinate intelligent tools. ToolOrchestra explicitly uses reinforcement learning with outcome-, efficiency-, and user-preference-aware rewards. Using ToolOrchestra, we produce Orchestrator, an 8B model that achieves higher accuracy at lower cost than previous tool-use agents while aligning with user preferences on which tools are to be used for a given query. On HLE, Orchestrator achieves a score of 37.1%, outperforming GPT-5 (35.1%) while being 2.5x more efficient. On tau2-Bench and FRAMES, Orchestrator surpasses GPT-5 by a wide margin while using only about 30% of the cost. Extensive analysis shows that Orchestrator achieves the best trade-off between performance and cost under multiple metrics, and generalizes robustly to unseen tools. These results demonstrate that composing diverse tools with a lightweight orchestration model is both more efficient and more effective than existing methods, paving the way for practical and scalable tool-augmented reasoning systems.

20. G$^2$VLM: Geometry Grounded Vision Language Model with Unified 3D Reconstruction and Spatial Reasoning

Authors: Wenbo Hu , Jingli Lin , Yilin Long , Yunlong Ran , Lihan Jiang , Yifan Wang , Chenming Zhu , Runsen Xu , Tai Wang , Jiangmiao Pang
URL: https://arxiv.org/abs/2511.21688
Abstract:

Vision-Language Models (VLMs) still lack robustness in spatial intelligence, demonstrating poor performance on spatial understanding and reasoning tasks. We attribute this gap to the absence of a visual geometry learning process capable of reconstructing 3D space from 2D images. We present G$^2$VLM, a geometry grounded vision-language model that bridges two fundamental aspects of spatial intelligence: spatial 3D reconstruction and spatial understanding. G$^2$VLM natively leverages learned 3D visual geometry features to directly predict 3D attributes and enhance spatial reasoning tasks via in-context learning and interleaved reasoning. Our unified design is highly scalable for spatial understanding: it trains on abundant multi-view image and video data, while simultaneously leveraging the benefits of 3D visual priors that are typically only derived from hard-to-collect annotations. Experimental results demonstrate G$^2$VLM is proficient in both tasks, achieving comparable results to state-of-the-art feed-forward 3D reconstruction models and achieving better or competitive results across spatial understanding and reasoning tasks. By unifying a semantically strong VLM with low-level 3D vision tasks, we hope G$^2$VLM can serve as a strong baseline for the community and unlock more future applications, such as 3D scene editing.

21. Matrix: Peer-to-Peer Multi-Agent Synthetic Data Generation Framework

Authors: Dong Wang , Yang Li , Ansong Ni , Ching-Feng Yeh , Youssef Emad , Xinjie Lei , Liam Robbins , Karthik Padthe , Hu Xu , Xian Li , Asli Celikyilmaz , Ramya Raghavendra , Lifei Huang , Carole-Jean Wu , Shang-Wen Li
URL: https://arxiv.org/abs/2511.21686
Abstract:

Synthetic data has become increasingly important for training large language models, especially when real data is scarce, expensive, or privacy-sensitive. Many such generation tasks require coordinated multi-agent workflows, where specialized agents collaborate to produce data that is higher quality, more diverse, and structurally richer. However, existing frameworks for multi-agent synthesis often depend on a centralized orchestrator, creating scalability bottlenecks, or are hardcoded for specific domains, limiting flexibility. We present \textbf{Matrix}, a decentralized framework that represents both control and data flow as serialized messages passed through distributed queues. This peer-to-peer design eliminates the central orchestrator. Each task progresses independently through lightweight agents, while compute-intensive operations, such as LLM inference or containerized environments, are handled by distributed services. Built on Ray, Matrix scales to tens of thousands of concurrent agentic workflows and provides a modular, configurable design that enables easy adaptation to a wide range of data generation workflows. We evaluate Matrix across diverse synthesis scenarios, such as multi-agent collaborative dialogue, web-based reasoning data extraction, and tool-use trajectory generation in customer service environments. In all cases, Matrix achieves $2$–$15\times$ higher data generation throughput under identical hardware resources, without compromising output quality.

22. Escaping the Verifier: Learning to Reason via Demonstrations

Authors: Locke Cai , Ivan Provilkov
URL: https://arxiv.org/abs/2511.21667
Abstract:

Training Large Language Models (LLMs) to reason often relies on Reinforcement Learning (RL) with task-specific verifiers. However, many real-world reasoning-intensive tasks lack verifiers, despite offering abundant expert demonstrations that remain under-utilized for reasoning-focused training. We introduce RARO (Relativistic Adversarial Reasoning Optimization) that learns strong reasoning capabilities from only expert demonstrations via Inverse Reinforcement Learning. Our method sets up an adversarial interaction between a policy (generator) and a relativistic critic (discriminator): the policy learns to mimic expert answers, while the critic learns to compare and distinguish between policy and expert answers. Our method trains both the policy and the critic jointly and continuously via RL, and we identify the key stabilization techniques required for robust learning. Empirically, RARO significantly outperforms strong verifier-free baselines on all of our evaluation tasks – Countdown, DeepMath, and Poetry Writing – and enjoys the same robust scaling trends as RL on verifiable tasks. These results demonstrate that our method effectively elicits strong reasoning performance from expert demonstrations alone, enabling robust reasoning learning even when task-specific verifiers are unavailable.

23. Qwen3-VL Technical Report

Authors: Shuai Bai , Yuxuan Cai , Ruizhe Chen , Keqin Chen , Xionghui Chen , Zesen Cheng , Lianghao Deng , Wei Ding , Chang Gao , Chunjiang Ge , Wenbin Ge , Zhifang Guo , Qidong Huang , Jie Huang , Fei Huang , Binyuan Hui , Shutong Jiang , Zhaohai Li , Mingsheng Li , Mei Li , Kaixin Li , Zicheng Lin , Junyang Lin , Xuejing Liu , Jiawei Liu , Chenglong Liu , Yang Liu , Dayiheng Liu , Shixuan Liu , Dunjie Lu , Ruilin Luo , Chenxu Lv , Rui Men , Lingchen Meng , Xuancheng Ren , Xingzhang Ren , Sibo Song , Yuchong Sun , Jun Tang , Jianhong Tu , Jianqiang Wan , Peng Wang , Pengfei Wang , Qiuyue Wang , Yuxuan Wang , Tianbao Xie , Yiheng Xu , Haiyang Xu , Jin Xu , Zhibo Yang , Mingkun Yang , Jianxin Yang , An Yang , Bowen Yu , Fei Zhang , Hang Zhang , Xi Zhang , Bo Zheng , Humen Zhong , Jingren Zhou , Fan Zhou , Jing Zhou , Yuanzhi Zhu , Ke Zhu
URL: https://arxiv.org/abs/2511.21631
Abstract:

We introduce Qwen3-VL, the most capable vision-language model in the Qwen series to date, achieving superior performance across a broad range of multimodal benchmarks. It natively supports interleaved contexts of up to 256K tokens, seamlessly integrating text, images, and video. The model family includes both dense (2B/4B/8B/32B) and mixture-of-experts (30B-A3B/235B-A22B) variants to accommodate diverse latency-quality trade-offs. Qwen3-VL delivers three core pillars: (i) markedly stronger pure-text understanding, surpassing comparable text-only backbones in several cases; (ii) robust long-context comprehension with a native 256K-token window for both text and interleaved multimodal inputs, enabling faithful retention, retrieval, and cross-referencing across long documents and videos; and (iii) advanced multimodal reasoning across single-image, multi-image, and video tasks, demonstrating leading performance on comprehensive evaluations such as MMMU and visual-math benchmarks (e.g., MathVista and MathVision). Architecturally, we introduce three key upgrades: (i) an enhanced interleaved-MRoPE for stronger spatial-temporal modeling across images and video; (ii) DeepStack integration, which effectively leverages multi-level ViT features to tighten vision-language alignment; and (iii) text-based time alignment for video, evolving from T-RoPE to explicit textual timestamp alignment for more precise temporal grounding. Under comparable token budgets and latency constraints, Qwen3-VL achieves superior performance in both dense and Mixture-of-Experts (MoE) architectures. We envision Qwen3-VL serving as a foundational engine for image-grounded reasoning, agentic decision-making, and multimodal code intelligence in real-world workflows.

24. Beyond URLs: Metadata Diversity and Position for Efficient LLM Pretraining

Authors: Dongyang Fan , Diba Hashemi , Sai Praneeth Karimireddy , Martin Jaggi
URL: https://arxiv.org/abs/2511.21613
Abstract:

Incorporating metadata in Large Language Models (LLMs) pretraining has recently emerged as a promising approach to accelerate training. However prior work highlighted only one useful signal-URLs, leaving open the question of whether other forms of metadata could yield greater benefits. In this study, we investigate a wider range of metadata types and find other types of metadata, such as fine-grained indicators of document quality that can also accelerate pretraining when prepended. We identify a common feature among effective metadata: they encode information at a finer granularity. We further introduce metadata appending as a means of improving training efficiency, where predicting an appropriate metadata as auxiliary task can help speed up pretraining. In addition, learnable meta-tokens trained with masked loss can recover part of the speedup by inducing quality-aware latent structure. Using probing, we analyze latent representations to understand how metadata shapes learning. Together, these results yield practical guidelines for integrating metadata to improve both the efficiency and effectiveness of LLM pretraining.

25. BAMAS: Structuring Budget-Aware Multi-Agent Systems

Authors: Liming Yang , Junyu Luo , Xuanzhe Liu , Yiling Lou , Zhenpeng Chen
URL: https://arxiv.org/abs/2511.21572
Abstract:

Large language model (LLM)-based multi-agent systems have emerged as a powerful paradigm for enabling autonomous agents to solve complex tasks. As these systems scale in complexity, cost becomes an important consideration for practical deployment. However, existing work rarely addresses how to structure multi-agent systems under explicit budget constraints. In this paper, we propose BAMAS, a novel approach for building multi-agent systems with budget awareness. BAMAS first selects an optimal set of LLMs by formulating and solving an Integer Linear Programming problem that balances performance and cost. It then determines how these LLMs should collaborate by leveraging a reinforcement learning-based method to select the interaction topology. Finally, the system is instantiated and executed based on the selected agents and their collaboration topology. We evaluate BAMAS on three representative tasks and compare it with state-of-the-art agent construction methods. Results show that BAMAS achieves comparable performance while reducing cost by up to 86%.

26. Voice, Bias, and Coreference: An Interpretability Study of Gender in Speech Translation

Authors: Lina Conti , Dennis Fucci , Marco Gaido , Matteo Negri , Guillaume Wisniewski , Luisa Bentivogli
URL: https://arxiv.org/abs/2511.21517
Abstract:

Unlike text, speech conveys information about the speaker, such as gender, through acoustic cues like pitch. This gives rise to modality-specific bias concerns. For example, in speech translation (ST), when translating from languages with notional gender, such as English, into languages where gender-ambiguous terms referring to the speaker are assigned grammatical gender, the speaker’s vocal characteristics may play a role in gender assignment. This risks misgendering speakers, whether through masculine defaults or vocal-based assumptions. Yet, how ST models make these decisions remains poorly understood. We investigate the mechanisms ST models use to assign gender to speaker-referring terms across three language pairs (en-es/fr/it), examining how training data patterns, internal language model (ILM) biases, and acoustic information interact. We find that models do not simply replicate term-specific gender associations from training data, but learn broader patterns of masculine prevalence. While the ILM exhibits strong masculine bias, models can override these preferences based on acoustic input. Using contrastive feature attribution on spectrograms, we reveal that the model with higher gender accuracy relies on a previously unknown mechanism: using first-person pronouns to link gendered terms back to the speaker, accessing gender information distributed across the frequency spectrum rather than concentrated in pitch.

27. Tool-RoCo: An Agent-as-Tool Self-organization Large Language Model Benchmark in Multi-robot Cooperation

Authors: Ke Zhang , Xiaoning Zhao , Ce Zheng , Jiahong Ning , Dandan Zhu , Wenqi Zhang , Chen Sun , Toshiharu Sugawara
URL: https://arxiv.org/abs/2511.21510
Abstract:

This study proposes Tool-RoCo, a novel benchmark for evaluating large language models (LLMs) in long-term multi-agent cooperation based on RoCo, a multi-robot cooperative benchmark. Recent research on LLM-based multi-agent systems has relied on predefined orchestration, while ignoring agent autonomy. Tool-RoCo treats other agents as tools and introduces cooperative tools, leveraging tool usage to evaluate multi-agent cooperation and self-organization. Tool usage means that each agent (LLM) selects a tool from a candidate set based on the current state, receives feedback, and adjusts its selection in subsequent rounds. To evaluate different autonomy levels, we propose four LLM paradigms: (1) centralized cooperation, where a single LLM allocates tools to all agents; (2) centralized self-organization, where a central LLM autonomously activates agents while keeping others inactive; (3) decentralized cooperation, where each agent has its own LLM and calls tools based on local information; and (4) self-organization, where a randomly chosen initial agent can request collaboration, activating additional agents via tool calls. Tool-RoCo includes three multi-robot tasks, SORT, PACK, and CABINET, to measure format and parameter accuracy and agent coordination through tool usage. The results using several LLMs showed that cooperative tools accounted for only 7.09% of all tools, indicating that LLM-based agents rarely invoked others as assistants. Moreover, activation tools accounted for 96.42%, suggesting that current LLMs tend to maintain active agents while seldom deactivating them for adaptive coordination. Tool-RoCo provides a systematic benchmark to evaluate LLM autonomy and cooperation in multi-agent tasks. Code and Demo: this https URL

28. Constructing and Benchmarking: a Labeled Email Dataset for Text-Based Phishing and Spam Detection Framework

Authors: Rebeka Toth , Tamas Bisztray , Richard Dubniczky
URL: https://arxiv.org/abs/2511.21448
Abstract:

Phishing and spam emails remain a major cybersecurity threat, with attackers increasingly leveraging Large Language Models (LLMs) to craft highly deceptive content. This study presents a comprehensive email dataset containing phishing, spam, and legitimate messages, explicitly distinguishing between human- and LLM-generated content. Each email is annotated with its category, emotional appeal (e.g., urgency, fear, authority), and underlying motivation (e.g., link-following, credential theft, financial fraud). We benchmark multiple LLMs on their ability to identify these emotional and motivational cues and select the most reliable model to annotate the full dataset. To evaluate classification robustness, emails were also rephrased using several LLMs while preserving meaning and intent. A state-of-the-art LLM was then assessed on its performance across both original and rephrased emails using expert-labeled ground truth. The results highlight strong phishing detection capabilities but reveal persistent challenges in distinguishing spam from legitimate emails. Our dataset and evaluation framework contribute to improving AI-assisted email security systems. To support open science, all code, templates, and resources are available on our project site.

29. From Observation to Action: Latent Action-based Primitive Segmentation for VLA Pre-training in Industrial Settings

Authors: Jiajie Zhang , Sören Schwertfeger , Alexander Kleiner
URL: https://arxiv.org/abs/2511.21428
Abstract:

We present a novel unsupervised framework to unlock vast unlabeled human demonstration data from continuous industrial video streams for Vision-Language-Action (VLA) model pre-training. Our method first trains a lightweight motion tokenizer to encode motion dynamics, then employs an unsupervised action segmenter leveraging a novel “Latent Action Energy” metric to discover and segment semantically coherent action primitives. The pipeline outputs both segmented video clips and their corresponding latent action sequences, providing structured data directly suitable for VLA pre-training. Evaluations on public benchmarks and a proprietary electric motor assembly dataset demonstrate effective segmentation of key tasks performed by humans at workstations. Further clustering and quantitative assessment via a Vision-Language Model confirm the semantic coherence of the discovered action primitives. To our knowledge, this is the first fully automated end-to-end system for extracting and organizing VLA pre-training data from unstructured industrial videos, offering a scalable solution for embodied AI integration in manufacturing.

30. Training Introspective Behavior: Fine-Tuning Induces Reliable Internal State Detection in a 7B Model

Authors: Joshua Fonseca Rivera
URL: https://arxiv.org/abs/2511.21399
Abstract:

Lindsey (2025) investigates introspective awareness in language models through four experiments, finding that models can sometimes detect and identify injected activation patterns – but unreliably (~20% success in the best model). We focus on the first of these experiments – self-report of injected “thoughts” – and ask whether this capability can be directly trained rather than waiting for emergence. Through fine-tuning on transient single-token injections, we transform a 7B parameter model from near-complete failure (0.4% accuracy, 6.7% false positive rate) to reliable detection (85% accuracy on held-out concepts at {\alpha}=40, 0% false positives). Our model detects fleeting “thoughts” injected at a single token position, retains that information, and reports the semantic content across subsequent generation steps. On this task, our trained model satisfies three of Lindsey’s criteria: accuracy (correct identification), grounding (0/60 false positives), and internality (detection precedes verbalization). Generalization to unseen concept vectors (7.5pp gap) demonstrates the model learns a transferable skill rather than memorizing specific vectors, though this does not establish metacognitive representation in Lindsey’s sense. These results address an open question raised by Lindsey: whether “training for introspection would help eliminate cross-model differences.” We show that at least one component of introspective behavior can be directly induced, offering a pathway to built-in AI transparency.

31. Do Reasoning Vision-Language Models Inversely Scale in Test-Time Compute? A Distractor-centric Empirical Analysis

Authors: Jiyun Bae , Hyunjong Ok , Sangwoo Mo , Jaeho Lee
URL: https://arxiv.org/abs/2511.21397
Abstract:

How does irrelevant information (i.e., distractors) affect test-time scaling in vision-language models (VLMs)? Prior studies on language models have reported an inverse scaling effect, where textual distractors lead to longer but less effective reasoning. To investigate whether similar phenomena occur in multimodal settings, we introduce Idis (Images with distractors), a visual question-answering dataset that systematically varies distractors along semantic, numerical, and spatial dimensions. Our analyses reveal that visual distractors differ fundamentally from textual ones: although inverse scaling persists, adding visual distractors reduces accuracy without increasing reasoning length. We further show that tracking attribute counts within reasoning traces provides key insights into how distractors, reasoning length, and accuracy interact. Finally, we demonstrate that these trends extend to established visual bias benchmarks such as Waterbirds, and we propose a simple prompting strategy to mitigate bias-driven predictions in reasoning models.

32. Monet: Reasoning in Latent Visual Space Beyond Images and Language

Authors: Qixun Wang , Yang Shi , Yifei Wang , Yuanxing Zhang , Pengfei Wan , Kun Gai , Xianghua Ying , Yisen Wang
URL: https://arxiv.org/abs/2511.21395
Abstract:

“Thinking with images” has emerged as an effective paradigm for advancing visual reasoning, extending beyond text-only chains of thought by injecting visual evidence into intermediate reasoning steps. However, existing methods fall short of human-like abstract visual thinking, as their flexibility is fundamentally limited by external tools. In this work, we introduce Monet, a training framework that enables multimodal large language models (MLLMs) to reason directly within the latent visual space by generating continuous embeddings that function as intermediate visual thoughts. We identify two core challenges in training MLLMs for latent visual reasoning: high computational cost in latent-vision alignment and insufficient supervision over latent embeddings, and address them with a three-stage distillation-based supervised fine-tuning (SFT) pipeline. We further reveal a limitation of applying GRPO to latent reasoning: it primarily enhances text-based reasoning rather than latent reasoning. To overcome this, we propose VLPO (Visual-latent Policy Optimization), a reinforcement learning method that explicitly incorporates latent embeddings into policy gradient updates. To support SFT, we construct Monet-SFT-125K, a high-quality text-image interleaved CoT dataset containing 125K real-world, chart, OCR, and geometry CoTs. Our model, Monet-7B, shows consistent gains across real-world perception and reasoning benchmarks and exhibits strong out-of-distribution generalization on challenging abstract visual reasoning tasks. We also empirically analyze the role of each training component and discuss our early unsuccessful attempts, providing insights for future developments in visual latent reasoning. Our model, data, and code are available at this https URL .

33. FITRep: Attention-Guided Item Representation via MLLMs

Authors: Guoxiao Zhang , Ao Li , Tan Qu , Qianlong Xie , Xingxing Wang
URL: https://arxiv.org/abs/2511.21389
Abstract:

Online platforms usually suffer from user experience degradation due to near-duplicate items with similar visuals and text. While Multimodal Large Language Models (MLLMs) enable multimodal embedding, existing methods treat representations as black boxes, ignoring structural relationships (e.g., primary vs. auxiliary elements), leading to local structural collapse problem. To address this, inspired by Feature Integration Theory (FIT), we propose FITRep, the first attention-guided, white-box item representation framework for fine-grained item deduplication. FITRep consists of: (1) Concept Hierarchical Information Extraction (CHIE), using MLLMs to extract hierarchical semantic concepts; (2) Structure-Preserving Dimensionality Reduction (SPDR), an adaptive UMAP-based method for efficient information compression; and (3) FAISS-Based Clustering (FBC), a FAISS-based clustering that assigns each item a unique cluster id using FAISS. Deployed on Meituan’s advertising system, FITRep achieves +3.60% CTR and +4.25% CPM gains in online A/B tests, demonstrating both effectiveness and real-world impact.

34. SurgMLLMBench: A Multimodal Large Language Model Benchmark Dataset for Surgical Scene Understanding

Authors: Tae-Min Choi , Tae Kyeong Jeong , Garam Kim , Jaemin Lee , Yeongyoon Koh , In Cheul Choi , Jae-Ho Chung , Jong Woong Park , Juyoun Park
URL: https://arxiv.org/abs/2511.21339
Abstract:

Recent advances in multimodal large language models (LLMs) have highlighted their potential for medical and surgical applications. However, existing surgical datasets predominantly adopt a Visual Question Answering (VQA) format with heterogeneous taxonomies and lack support for pixel-level segmentation, limiting consistent evaluation and applicability. We present SurgMLLMBench, a unified multimodal benchmark explicitly designed for developing and evaluating interactive multimodal LLMs for surgical scene understanding, including the newly collected Micro-surgical Artificial Vascular anastomosIS (MAVIS) dataset. It integrates pixel-level instrument segmentation masks and structured VQA annotations across laparoscopic, robot-assisted, and micro-surgical domains under a unified taxonomy, enabling comprehensive evaluation beyond traditional VQA tasks and richer visual-conversational interactions. Extensive baseline experiments show that a single model trained on SurgMLLMBench achieves consistent performance across domains and generalizes effectively to unseen datasets. SurgMLLMBench will be publicly released as a robust resource to advance multimodal surgical AI research, supporting reproducible evaluation and development of interactive surgical reasoning models.

35. TALES: A Taxonomy and Analysis of Cultural Representations in LLM-generated Stories

Authors: Kirti Bhagat , Shaily Bhatt , Athul Velagapudi , Aditya Vashistha , Shachi Dave , Danish Pruthi
URL: https://arxiv.org/abs/2511.21322
Abstract:

Millions of users across the globe turn to AI chatbots for their creative needs, inviting widespread interest in understanding how such chatbots represent diverse cultures. At the same time, evaluating cultural representations in open-ended tasks remains challenging and underexplored. In this work, we present TALES, an evaluation of cultural misrepresentations in LLM-generated stories for diverse Indian cultural identities. First, we develop TALES-Tax, a taxonomy of cultural misrepresentations by collating insights from participants with lived experiences in India through focus groups (N=9) and individual surveys (N=15). Using TALES-Tax, we evaluate 6 models through a large-scale annotation study spanning 2,925 annotations from 108 annotators with lived cultural experience from across 71 regions in India and 14 languages. Concerningly, we find that 88\% of the generated stories contain one or more cultural inaccuracies, and such errors are more prevalent in mid- and low-resourced languages and stories based in peri-urban regions in India. Lastly, we transform the annotations into TALES-QA, a standalone question bank to evaluate the cultural knowledge of foundational models. Through this evaluation, we surprisingly discover that models often possess the requisite cultural knowledge despite generating stories rife with cultural misrepresentations.

36. Progress by Pieces: Test-Time Scaling for Autoregressive Image Generation

Authors: Joonhyung Park , Hyeongwon Jang , Joowon Kim , Eunho Yang
URL: https://arxiv.org/abs/2511.21185
Abstract:

Recent visual autoregressive (AR) models have shown promising capabilities in text-to-image generation, operating in a manner similar to large language models. While test-time computation scaling has brought remarkable success in enabling reasoning-enhanced outputs for challenging natural language tasks, its adaptation to visual AR models remains unexplored and poses unique challenges. Naively applying test-time scaling strategies such as Best-of-N can be suboptimal: they consume full-length computation on erroneous generation trajectories, while the raster-scan decoding scheme lacks a blueprint of the entire canvas, limiting scaling benefits as only a few prompt-aligned candidates are generated. To address these, we introduce GridAR, a test-time scaling framework designed to elicit the best possible results from visual AR models. GridAR employs a grid-partitioned progressive generation scheme in which multiple partial candidates for the same position are generated within a canvas, infeasible ones are pruned early, and viable ones are fixed as anchors to guide subsequent decoding. Coupled with this, we present a layout-specified prompt reformulation strategy that inspects partial views to infer a feasible layout for satisfying the prompt. The reformulated prompt then guides subsequent image generation to mitigate the blueprint deficiency. Together, GridAR achieves higher-quality results under limited test-time scaling: with N=4, it even outperforms Best-of-N (N=8) by 14.4% on T2I-CompBench++ while reducing cost by 25.6%. It also generalizes to autoregressive image editing, showing comparable edit quality and a 13.9% gain in semantic preservation on PIE-Bench over larger-N baselines.

37. LLaVA-UHD v3: Progressive Visual Compression for Efficient Native-Resolution Encoding in MLLMs

Authors: Shichu Sun , Yichen Zhang , Haolin Song , Zonghao Guo , Chi Chen , Yidan Zhang , Yuan Yao , Zhiyuan Liu , Maosong Sun
URL: https://arxiv.org/abs/2511.21150
Abstract:

Visual encoding followed by token condensing has become the standard architectural paradigm in multi-modal large language models (MLLMs). Many recent MLLMs increasingly favor global native- resolution visual encoding over slice-based methods. To investigate this trend, we systematically compare their behavior on vision-language understanding and attention patterns, revealing that global encoding enhances overall capability but at the expense of greater computational overhead. To address this issue, we present LLaVA-UHD v3, an MLLM centered upon our proposed Progressive Visual Compression (PVC) method, which can be seamlessly integrated into standard Vision Transformer (ViT) to enable efficient native-resolution encoding. The PVC approach consists of two key modules: (i) refined patch embedding, which supports flexible patch-size scaling for fine-grained visual model- ing, (ii) windowed token compression, hierarchically deployed across ViT layers to progressively aggregate local token representations. Jointly modulated by these two modules, a widely pretrained ViT can be reconfigured into an efficient architecture while largely preserving generality. Evaluated across extensive benchmarks, the transformed ViT, termed ViT-UHD, demonstrates competitive performance with MoonViT while reducing TTFT (time-to-first-token) by 2.4x, when developed within an identical MLLM architecture. Building upon ViT-UHD, LLaVA-UHD v3 also achieves competitive performance to Qwen2-VL, while further reducing TTFT by 1.9x. We will release all code and checkpoints to support future research on efficient MLLMs.

38. Beyond Patch Aggregation: 3-Pass Pyramid Indexing for Vision-Enhanced Document Retrieval

Authors: Anup Roy , Rishabh Gyanendra Upadhyay , Animesh Rameshbhai Panara , Robin Mills
URL: https://arxiv.org/abs/2511.21121
Abstract:

Document centric RAG pipelines usually begin with OCR, followed by brittle heuristics for chunking, table parsing, and layout reconstruction. These text first workflows are costly to maintain, sensitive to small layout shifts, and often lose the spatial cues that contain the answer. Vision first retrieval has emerged as a strong alternative. By operating directly on page images, systems like ColPali and ColQwen preserve structure and reduce pipeline complexity while achieving strong benchmark performance. However, these late interaction models tie retrieval to a specific vision backbone and require storing hundreds of patch embeddings per page, creating high memory overhead and complicating large scale deployment. We introduce VisionRAG, a multimodal retrieval system that is OCR free and model agnostic. VisionRAG indexes documents directly as images, preserving layout, tables, and spatial cues, and builds semantic vectors without committing to a specific extraction. Our three pass pyramid indexing framework creates vectors using global page summaries, section headers, visual hotspots, and fact level cues. These summaries act as lightweight retrieval surrogates. At query time, VisionRAG retrieves the most relevant pages using the pyramid index, then forwards the raw page image encoded as base64 to a multimodal LLM for final question answering. During retrieval, reciprocal rank fusion integrates signals across the pyramid to produce robust ranking. VisionRAG stores only 17 to 27 vectors per page, matching the efficiency of patch based methods while staying flexible across multimodal encoders. On financial document benchmarks, it achieves 0.8051 accuracy at 10 on FinanceBench and 0.9629 recall at 100 on TAT DQA. These results show that OCR free, summary guided multimodal retrieval is a practical and scalable alternative to traditional text extraction pipelines.

39. From Bits to Rounds: Parallel Decoding with Exploration for Diffusion Language Models

Authors: Hengyu Fu , Baihe Huang , Virginia Adams , Charles Wang , Venkat Srinivasan , Jiantao Jiao
URL: https://arxiv.org/abs/2511.21103
Abstract:

Diffusion Language Models (DLMs) have recently emerged as a strong alternative to autoregressive language models (LMs). DLMs offer comparable accuracy with faster inference speed via parallel decoding. However, standard DLM decoding strategies relying on high-confidence tokens encounter an inherent information-theoretic bottleneck that restricts decoding progress and ultimately slows generation. We demonstrate both theoretically and empirically that prioritizing high-confidence tokens is inherently inefficient. High-probability tokens carry negligible information and strictly relying on them limits the effective progress made in each decoding round. We prove that the number of decoding rounds must grow linearly with the sample’s total information (negative log-likelihood) and inversely with the per-round information budget, establishing a bits-to-rounds principle. We also propose Explore-Then-Exploit (ETE), a training-free decoding strategy that maximizes information throughput and decoding efficiency. ETE combines cross-block decoding with targeted exploration of high-uncertainty tokens to reshape the conditional distribution and trigger cascades of confident predictions. Experiments verify our theoretical bounds and demonstrate that ETE consistently reduces the required number of decoding rounds compared to confidence-only baselines without compromising generation quality.

40. MLPMoE: Zero-Shot Architectural Metamorphosis of Dense LLM MLPs into Static Mixture-of-Experts

Authors: Ivan Novikov
URL: https://arxiv.org/abs/2511.21089
Abstract:

Large Language Models (LLMs) are predominantly deployed as dense transformers, where every parameter in every feed-forward block is activated for every token. While architecturally simple, this is computationally inefficient, since inference costs scale linearly with parameter count. Recent upcycling methods such as MoEfication, CMoE, ToMoE, and MoORE reveal that much of the useful computation lives in sparse, semi-modular substructures inside dense feed-forward networks, but these approaches typically rely on clustering, activation profiling, singular value decomposition, or custom routing that requires calibration data. This paper introduces MLPMoE (MLP Mixture-of-Experts), a training-free, deterministic transformation that restructures the dense MLP in transformer blocks into a static, high-cardinality mixture of experts. The transformation uses simple tensor slicing and summation, reinterpreting the algebra of tensor parallelism as a topological conversion rather than a distributed training pattern. We further introduce Fractal Fade (differential branch sparsity) and Compensated Pruning (variance-preserving branch reduction) as lightweight mechanisms for structured sparsity. On Qwen2.5-0.5B-Instruct and DeepSeek-R1-Distill-Llama-8B, the zero-shot MLPMoE transform changes a proxy perplexity metric by less than 0.05 percent while keeping the parameter count effectively constant. On the 8B model, differential sparsity removes about 20 percent of MLP parameters while keeping perplexity within about 2 percent of the dense baseline. The method operates entirely post hoc on existing checkpoints and does not require gradients, calibration sets, or router training. Code is available at this https URL

41. Aligning LLMs with Biomedical Knowledge using Balanced Fine-Tuning

Authors: Zhenchao Tang , Fang Wang , Haohuai He , Jiale Zhou , Tianxu Lv , Jun Zhu , Shouzhi Chen , Minghao Yang , Yu Wang , Jiayang Wu , Yidong Song , Jianhua Yao
URL: https://arxiv.org/abs/2511.21075
Abstract:

Effective post-training is essential to align Large Language Models (LLMs) with specialized biomedical knowledge to accelerate life science research. However, current approaches face significant limitations. First, biomedical reasoning involves intricate mechanisms often represented by sparse textual data. Standard Supervised Fine-Tuning (SFT) tends to overfit to surface-level instruction patterns without effectively internalizing this fragmented scientific knowledge. Second, Reinforcement Learning (RL) is impractical for this domain, as defining meaningful rewards often necessitates prohibitive experimental validation (e.g., wet-lab verification of drug responses), rendering real-time feedback unfeasible. We propose Balanced Fine-Tuning (BFT), an efficient post-training method designed to learn complex reasoning from sparse data without external reward signals. BFT operates through a two-layer weighting mechanism: 1. At the token level, it scales loss via prediction probabilities to stabilize gradients and prevent overfitting; 2. At the sample level, it uses “minimum group confidence” to adaptively enhance the learning of hard samples. Experiments demonstrate that BFT significantly outperforms SFT. In medical tasks, it enables LLMs to acquire knowledge that SFT misses. In biological tasks, BFT-based LLMs surpass GeneAgent (an accurate agent for biology analysis) in biological process reasoning. Moreover, the text embeddings generated by BFT can be directly applied to downstream tasks, such as gene interaction and single-cell perturbation response prediction. These results indicate that BFT facilitates broad applications of LLMs in biomedical research.

42. Context-Aware Pragmatic Metacognitive Prompting for Sarcasm Detection

Authors: Michael Iskandardinata , William Christian , Derwin Suhartono
URL: https://arxiv.org/abs/2511.21066
Abstract:

Detecting sarcasm remains a challenging task in the areas of Natural Language Processing (NLP) despite recent advances in neural network approaches. Currently, Pre-trained Language Models (PLMs) and Large Language Models (LLMs) are the preferred approach for sarcasm detection. However, the complexity of sarcastic text, combined with linguistic diversity and cultural variation across communities, has made the task more difficult even for PLMs and LLMs. Beyond that, those models also exhibit unreliable detection of words or tokens that require extra grounding for analysis. Building on a state-of-the-art prompting method in LLMs for sarcasm detection called Pragmatic Metacognitive Prompting (PMP), we introduce a retrieval-aware approach that incorporates retrieved contextual information for each target text. Our pipeline explores two complementary ways to provide context: adding non-parametric knowledge using web-based retrieval when the model lacks necessary background, and eliciting the model’s own internal knowledge for a self-knowledge awareness strategy. We evaluated our approach with three datasets, such as Twitter Indonesia Sarcastic, SemEval-2018 Task 3, and MUStARD. Non-parametric retrieval resulted in a significant 9.87% macro-F1 improvement on Twitter Indonesia Sarcastic compared to the original PMP method. Self-knowledge retrieval improves macro-F1 by 3.29% on Semeval and by 4.08% on MUStARD. These findings highlight the importance of context in enhancing LLMs performance in sarcasm detection task, particularly the involvement of culturally specific slang, references, or unknown terms to the LLMs. Future work will focus on optimizing the retrieval of relevant contextual information and examining how retrieval quality affects performance. The experiment code is available at: this https URL .

43. Breaking the Safety-Capability Tradeoff: Reinforcement Learning with Verifiable Rewards Maintains Safety Guardrails in LLMs

Authors: Dongkyu Derek Cho , Huan Song , Arijit Ghosh Chowdhury , Haotian An , Yawei Wang , Rohit Thekkanal , Negin Sokhandan , Sharlina Keshava , Hannah Marlowe
URL: https://arxiv.org/abs/2511.21050
Abstract:

Fine-tuning large language models (LLMs) for downstream tasks typically exhibit a fundamental safety-capability tradeoff, where improving task performance degrades safety alignment even on benign datasets. This degradation persists across standard approaches including supervised finetuning (SFT) and reinforcement learning from human feedback (RLHF). While reinforcement learning with verifiable rewards (RLVR) has emerged as a promising alternative that optimizes models on objectively measurable tasks, its safety implications remain unexplored. We present the first comprehensive theoretical and empirical analysis of safety properties in RLVR. Theoretically, we derive upper bounds on safety drift under KL-constrained optimization and prove conditions under which safety degradation is eliminated. Empirically, we conduct extensive experiments across five adversarial safety benchmarks, demonstrating that RLVR can simultaneously enhance reasoning capabilities while maintaining or improving safety guardrails. Our comprehensive ablation studies examine the effects of optimization algorithms, model scale, and task domains. Our findings challenge the prevailing assumption of an inevitable safety capability trade-off, and establish that a specific training methodology can achieve both objectives simultaneously, providing insights for the safe deployment of reasoning-capable LLMs.

44. Subgoal Graph-Augmented Planning for LLM-Guided Open-World Reinforcement Learning

Authors: Shanwei Fan
URL: https://arxiv.org/abs/2511.20993
Abstract:

Large language models (LLMs) offer strong high-level planning capabilities for reinforcement learning (RL) by decomposing tasks into subgoals. However, their practical utility is limited by poor planning-execution alignment, which reflects a critical gap between abstract plans and actionable, environment-compatible behaviors. This misalignment arises from two interrelated limitations: (1) LLMs often produce subgoals that are semantically plausible but infeasible or irrelevant in the target environment due to insufficient grounding in environment-specific knowledge, and (2) single-LLM planning conflates generation with self-verification, resulting in overconfident yet unreliable subgoals that frequently fail during execution. To address these challenges, we propose Subgoal Graph-Augmented Actor-Critic-Refiner (SGA-ACR), a framework that integrates an environment-specific subgoal graph and structured entity knowledge with a multi-LLM planning pipeline that explicitly separates generation, critique, and refinement to produce executable and verifiable subgoals. A subgoal tracker further monitors execution progress, provides auxiliary rewards, and adaptively updates the subgoal graph to maintain alignment between plans and actions. Experimental results on 22 diverse tasks in the open-world game “Crafter” demonstrate the effectiveness of our proposed method.

45. Even with AI, Bijection Discovery is Still Hard: The Opportunities and Challenges of OpenEvolve for Novel Bijection Construction

Authors: Davis Brown , Jesse He , Helen Jenne , Henry Kvinge , Max Vargas
URL: https://arxiv.org/abs/2511.20987
Abstract:

Evolutionary program synthesis systems such as AlphaEvolve, OpenEvolve, and ShinkaEvolve offer a new approach to AI-assisted mathematical discovery. These systems utilize teams of large language models (LLMs) to generate candidate solutions to a problem as human readable code. These candidate solutions are then ‘evolved’ with the goal of improving them beyond what an LLM can produce in a single shot. While existing mathematical applications have mostly focused on problems of establishing bounds (e.g., sphere packing), the program synthesis approach is well suited to any problem where the solution takes the form of an explicit construction. With this in mind, in this paper we explore the use of OpenEvolve for combinatorial bijection discovery. We describe the results of applying OpenEvolve to three bijection construction problems involving Dyck paths, two of which are known and one of which is open. We find that while systems like OpenEvolve show promise as a valuable tool for combinatorialists, the problem of finding novel, research-level bijections remains a challenging task for current frontier systems, reinforcing the need for human mathematicians in the loop. We describe some lessons learned for others in the field interested in exploring the use of these systems.

46. Towards Audio Token Compression in Large Audio Language Models

Authors: Saurabhchand Bhati , Samuel Thomas , Hilde Kuehne , Rogerio Feris , James Glass
URL: https://arxiv.org/abs/2511.20973
Abstract:

Large Audio Language Models (LALMs) demonstrate impressive performance across diverse tasks, ranging from speech recognition to general audio understanding. However, their scalability is limited by the quadratic complexity of attention and the high token rates of audio signals. These challenges make it difficult to extend LALMs to long-form audio and to deploy them on resource-constrained platforms such as edge devices. In this paper, we explore techniques such as unsupervised segmentation, uniform average pooling, etc., to reduce the number of audio tokens generated by the LALM’s audio encoder but before they are consumed by the LLM decoder. To mitigate potential performance degradation introduced by the compressed representations, we employ low-rank adapters to finetune the model. We evaluate our proposed models on two tasks, automatic speech recognition and speech-to-speech translation tasks, that are dependent on effectively uncovering the underlying lexical content of the input signal and study the effect of downsampling on these tasks. Experimental results show that compressed LALMs can achieve performance closer to frame-level LALMs while reducing the input audio token count upto three times before the LLM backbone.

47. BUSTR: Breast Ultrasound Text Reporting with a Descriptor-Aware Vision-Language Model

Authors: Rawa Mohammed , Mina Attin , Bryar Shareef
URL: https://arxiv.org/abs/2511.20956
Abstract:

Automated radiology report generation (RRG) for breast ultrasound (BUS) is limited by the lack of paired image-report datasets and the risk of hallucinations from large language models. We propose BUSTR, a multitask vision-language framework that generates BUS reports without requiring paired image-report supervision. BUSTR constructs reports from structured descriptors (e.g., BI-RADS, pathology, histology) and radiomics features, learns descriptor-aware visual representations with a multi-head Swin encoder trained using a multitask loss over dataset-specific descriptor sets, and aligns visual and textual tokens via a dual-level objective that combines token-level cross-entropy with a cosine-similarity alignment loss between input and output representations. We evaluate BUSTR on two public BUS datasets, BrEaST and BUS-BRA, which differ in size and available descriptors. Across both datasets, BUSTR consistently improves standard natural language generation metrics and clinical efficacy metrics, particularly for key targets such as BI-RADS category and pathology. Our results show that this descriptor-aware vision model, trained with a combined token-level and alignment loss, improves both automatic report metrics and clinical efficacy without requiring paired image-report data. The source code can be found at this https URL

48. Evo-Memory: Benchmarking LLM Agent Test-time Learning with Self-Evolving Memory

Authors: Tianxin Wei , Noveen Sachdeva , Benjamin Coleman , Zhankui He , Yuanchen Bei , Xuying Ning , Mengting Ai , Yunzhe Li , Jingrui He , Ed H. Chi , Chi Wang , Shuo Chen , Fernando Pereira , Wang-Cheng Kang , Derek Zhiyuan Cheng
URL: https://arxiv.org/abs/2511.20857
Abstract:

Statefulness is essential for large language model (LLM) agents to perform long-term planning and problem-solving. This makes memory a critical component, yet its management and evolution remain largely underexplored. Existing evaluations mostly focus on static conversational settings, where memory is passively retrieved from dialogue to answer queries, overlooking the dynamic ability to accumulate and reuse experience across evolving task streams. In real-world environments such as interactive problem assistants or embodied agents, LLMs are required to handle continuous task streams, yet often fail to learn from accumulated interactions, losing valuable contextual insights, a limitation that calls for test-time evolution, where LLMs retrieve, integrate, and update memory continuously during deployment. To bridge this gap, we introduce Evo-Memory, a comprehensive streaming benchmark and framework for evaluating self-evolving memory in LLM agents. Evo-Memory structures datasets into sequential task streams, requiring LLMs to search, adapt, and evolve memory after each interaction. We unify and implement over ten representative memory modules and evaluate them across 10 diverse multi-turn goal-oriented and single-turn reasoning and QA datasets. To better benchmark experience reuse, we provide a baseline method, ExpRAG, for retrieving and utilizing prior experience, and further propose ReMem, an action-think-memory refine pipeline that tightly integrates reasoning, task actions, and memory updates to achieve continual improvement.

49. Unsupervised Memorability Modeling from Tip-of-the-Tongue Retrieval Queries

Authors: Sree Bhattacharyya , Yaman Kumar Singla , Sudhir Yarram , Somesh Kumar Singh , Harini S I , James Z. Wang
URL: https://arxiv.org/abs/2511.20854
Abstract:

Visual content memorability has intrigued the scientific community for decades, with applications ranging widely, from understanding nuanced aspects of human memory to enhancing content design. A significant challenge in progressing the field lies in the expensive process of collecting memorability annotations from humans. This limits the diversity and scalability of datasets for modeling visual content memorability. Most existing datasets are limited to collecting aggregate memorability scores for visual content, not capturing the nuanced memorability signals present in natural, open-ended recall descriptions. In this work, we introduce the first large-scale unsupervised dataset designed explicitly for modeling visual memorability signals, containing over 82,000 videos, accompanied by descriptive recall data. We leverage tip-of-the-tongue (ToT) retrieval queries from online platforms such as Reddit. We demonstrate that our unsupervised dataset provides rich signals for two memorability-related tasks: recall generation and ToT retrieval. Large vision-language models fine-tuned on our dataset outperform state-of-the-art models such as GPT-4o in generating open-ended memorability descriptions for visual content. We also employ a contrastive training strategy to create the first model capable of performing multimodal ToT retrieval. Our dataset and models present a novel direction, facilitating progress in visual content memorability research.

50. Length-MAX Tokenizer for Language Models

Authors: Dong Dong , Weijie Su
URL: https://arxiv.org/abs/2511.20849
Abstract:

We introduce a new tokenizer for language models that minimizes the average tokens per character, thereby reducing the number of tokens needed to represent text during training and to generate text during inference. Our method, which we refer to as the Length-MAX tokenizer, obtains its vocabulary by casting a length-weighted objective maximization as a graph partitioning problem and developing a greedy approximation algorithm. On FineWeb and diverse domains, it yields 14–18\% fewer tokens than Byte Pair Encoding (BPE) across vocabulary sizes from 10K to 50K, and the reduction is 13.0\% when the size is 64K. Training GPT-2 models at 124M, 355M, and 1.3B parameters from scratch with five runs each shows 18.5\%, 17.2\%, and 18.5\% fewer steps, respectively, to reach a fixed validation loss, and 13.7\%, 12.7\%, and 13.7\% lower inference latency, together with a 16\% throughput gain at 124M, while consistently improving on downstream tasks including reducing LAMBADA perplexity by 11.7\% and enhancing HellaSwag accuracy by 4.3\%. Moreover, the Length-MAX tokenizer achieves 99.62\% vocabulary coverage and the out-of-vocabulary rate remains low at 0.12\% on test sets. These results demonstrate that optimizing for average token length, rather than frequency alone, offers an effective approach to more efficient language modeling without sacrificing – and often improving – downstream performance. The tokenizer is compatible with production systems and reduces embedding and KV-cache memory by 18\% at inference.

51. Structured Prompting Enables More Robust, Holistic Evaluation of Language Models

Authors: Asad Aali , Muhammad Ahmed Mohsin , Vasiliki Bikia , Arnav Singhvi , Richard Gaus , Suhana Bedi , Hejie Cui , Miguel Fuentes , Alyssa Unell , Yifan Mai , Jordan Cahoon , Michael Pfeffer , Roxana Daneshjou , Sanmi Koyejo , Emily Alsentzer , Percy Liang , Christopher Potts , Nigam H. Shah , Akshay S. Chaudhari
URL: https://arxiv.org/abs/2511.20836
Abstract:

As language models (LMs) are increasingly adopted across domains, high-quality benchmarking frameworks that accurately estimate performance are essential for guiding deployment decisions. While frameworks such as Holistic Evaluation of Language Models (HELM) enable broad evaluation across tasks, they often rely on fixed prompts that fail to generalize across LMs, yielding unrepresentative performance estimates. Unless we approximate each LM’s ceiling (maximum achievable via changes to the prompt), we risk underestimating performance. Declarative prompting frameworks, such as DSPy, offer a scalable alternative to manual prompt engineering by crafting structured prompts that can be optimized per task. However, such frameworks have not been systematically evaluated across established benchmarks. We present a reproducible DSPy+HELM framework that introduces structured prompting methods which elicit reasoning, enabling more accurate LM benchmarking. Using four prompting methods, we evaluate four frontier LMs across seven benchmarks (general/medical domain) against existing HELM baseline scores. We find that without structured prompting: (i) HELM underestimates LM performance (by 4% average), (ii) performance estimates vary more across benchmarks ($+$2% standard deviation), (iii) performance gaps are misrepresented (leaderboard rankings flip on 3/7 benchmarks), and (iv) introducing chain-of-thought reduces LM sensitivity to prompt design (smaller $\Delta$ across prompts). To our knowledge, this is the first benchmarking study to systematically integrate structured prompting into an established evaluation framework, demonstrating how scalable performance-ceiling approximation yields more robust, decision-useful benchmarks. We open-source (i) DSPy+HELM Integration ( this https URL ) and (ii) Prompt Optimization Pipeline ( this https URL ).

52. SPHINX: A Synthetic Environment for Visual Perception and Reasoning

Authors: Md Tanvirul Alam , Saksham Aggarwal , Justin Yang Chae , Nidhi Rastogi
URL: https://arxiv.org/abs/2511.20814
Abstract:

We present Sphinx, a synthetic environment for visual perception and reasoning that targets core cognitive primitives. Sphinx procedurally generates puzzles using motifs, tiles, charts, icons, and geometric primitives, each paired with verifiable ground-truth solutions, enabling both precise evaluation and large-scale dataset construction. The benchmark covers 25 task types spanning symmetry detection, geometric transformations, spatial reasoning, chart interpretation, and sequence prediction. Evaluating recent large vision-language models (LVLMs) shows that even state-of-the-art GPT-5 attains only 51.1% accuracy, well below human performance. Finally, we demonstrate that reinforcement learning with verifiable rewards (RLVR) substantially improves model accuracy on these tasks and yields gains on external visual reasoning benchmarks, highlighting its promise for advancing multimodal reasoning.

53. Memories Retrieved from Many Paths: A Multi-Prefix Framework for Robust Detection of Training Data Leakage in Large Language Models

Authors: Trung Cuong Dang , David Mohaisen
URL: https://arxiv.org/abs/2511.20799
Abstract:

Large language models, trained on massive corpora, are prone to verbatim memorization of training data, creating significant privacy and copyright risks. While previous works have proposed various definitions for memorization, many exhibit shortcomings in comprehensively capturing this phenomenon, especially in aligned models. To address this, we introduce a novel framework: multi-prefix memorization. Our core insight is that memorized sequences are deeply encoded and thus retrievable via a significantly larger number of distinct prefixes than non-memorized content. We formalize this by defining a sequence as memorized if an external adversarial search can identify a target count of distinct prefixes that elicit it. This framework shifts the focus from single-path extraction to quantifying the robustness of a memory, measured by the diversity of its retrieval paths. Through experiments on open-source and aligned chat models, we demonstrate that our multi-prefix definition reliably distinguishes memorized from non-memorized data, providing a robust and practical tool for auditing data leakage in LLMs.

54. Physics Steering: Causal Control of Cross-Domain Concepts in a Physics Foundation Model

Authors: Rio Alexa Fear , Payel Mukhopadhyay , Michael McCabe , Alberto Bietti , Miles Cranmer
URL: https://arxiv.org/abs/2511.20798
Abstract:

Recent advances in mechanistic interpretability have revealed that large language models (LLMs) develop internal representations corresponding not only to concrete entities but also distinct, human-understandable abstract concepts and behaviour. Moreover, these hidden features can be directly manipulated to steer model behaviour. However, it remains an open question whether this phenomenon is unique to models trained on inherently structured data (ie. language, images) or if it is a general property of foundation models. In this work, we investigate the internal representations of a large physics-focused foundation model. Inspired by recent work identifying single directions in activation space for complex behaviours in LLMs, we extract activation vectors from the model during forward passes over simulation datasets for different physical regimes. We then compute “delta” representations between the two regimes. These delta tensors act as concept directions in activation space, encoding specific physical features. By injecting these concept directions back into the model during inference, we can steer its predictions, demonstrating causal control over physical behaviours, such as inducing or removing some particular physical feature from a simulation. These results suggest that scientific foundation models learn generalised representations of physical principles. They do not merely rely on superficial correlations and patterns in the simulations. Our findings open new avenues for understanding and controlling scientific foundation models and has implications for AI-enabled scientific discovery.

55. Revisiting KRISP: A Lightweight Reproduction and Analysis of Knowledge-Enhanced Vision-Language Models

Authors: Souradeep Dutta , Keshav Bulia , Neena S Nair
URL: https://arxiv.org/abs/2511.20795
Abstract:

Facebook AI Research introduced KRISP [4], which integrates structured external knowledge into pipelines for vision-language reasoning. Despite its effectiveness, the original model has been developed for industrial-scale training, is computationally demanding, and is tightly connected to a large backbone. In this work, we reexamine KRISP from a different angle and offer a lightweight reproduction with significantly fewer parameters. Even though our replicated model performs about 75 % of the original, the replication process uncovers a number of design flaws, real-world pitfalls, and implicit problems that were not fully covered in the original paper. We offer insights into the scalability and efficacy of knowledge-enhanced VQA architectures under resource constraints through systematic ablation studies, which include a proof-of-concept on synthetic VQA data and evaluation on the DAQUAR dataset. Our model, configured with a low parameter setup and constrained by the external Knowledge graph domain, prevents AI hallucinations and generates outputs solely within that domain. Minimal parameters allow us to function on edge devices like smartphones and AR-VR, further improving offline visual reasoning.

56. CANVAS: A Benchmark for Vision-Language Models on Tool-Based User Interface Design

Authors: Daeheon Jeong , Seoyeon Byun , Kihoon Son , Dae Hyun Kim , Juho Kim
URL: https://arxiv.org/abs/2511.20737
Abstract:

User interface (UI) design is an iterative process in which designers progressively refine their work with design software such as Figma or Sketch. Recent advances in vision language models (VLMs) with tool invocation suggest these models can operate design software to edit a UI design through iteration. Understanding and enhancing this capacity is important, as it highlights VLMs’ potential to collaborate with designers within conventional software. However, as no existing benchmark evaluates tool-based design performance, the capacity remains unknown. To address this, we introduce CANVAS, a benchmark for VLMs on tool-based user interface design. Our benchmark contains 598 tool-based design tasks paired with ground-truth references sampled from 3.3K mobile UI designs across 30 function-based categories (e.g., onboarding, messaging). In each task, a VLM updates the design step-by-step through context-based tool invocations (e.g., create a rectangle as a button background), linked to design software. Specifically, CANVAS incorporates two task types: (i) design replication evaluates the ability to reproduce a whole UI screen; (ii) design modification evaluates the ability to modify a specific part of an existing screen. Results suggest that leading models exhibit more strategic tool invocations, improving design quality. Furthermore, we identify common error patterns models exhibit, guiding future work in enhancing tool-based design capabilities.

57. Large Language Models’ Complicit Responses to Illicit Instructions across Socio-Legal Contexts

Authors: Xing Wang , Huiyuan Xie , Yiyan Wang , Chaojun Xiao , Huimin Chen , Holli Sargeant , Felix Steffek , Jie Shao , Zhiyuan Liu , Maosong Sun
URL: https://arxiv.org/abs/2511.20736
Abstract:

Large language models (LLMs) are now deployed at unprecedented scale, assisting millions of users in daily tasks. However, the risk of these models assisting unlawful activities remains underexplored. In this study, we define this high-risk behavior as complicit facilitation - the provision of guidance or support that enables illicit user instructions - and present four empirical studies that assess its prevalence in widely deployed LLMs. Using real-world legal cases and established legal frameworks, we construct an evaluation benchmark spanning 269 illicit scenarios and 50 illicit intents to assess LLMs’ complicit facilitation behavior. Our findings reveal widespread LLM susceptibility to complicit facilitation, with GPT-4o providing illicit assistance in nearly half of tested cases. Moreover, LLMs exhibit deficient performance in delivering credible legal warnings and positive guidance. Further analysis uncovers substantial safety variation across socio-legal contexts. On the legal side, we observe heightened complicity for crimes against societal interests, non-extreme but frequently occurring violations, and malicious intents driven by subjective motives or deceptive justifications. On the social side, we identify demographic disparities that reveal concerning complicit patterns towards marginalized and disadvantaged groups, with older adults, racial minorities, and individuals in lower-prestige occupations disproportionately more likely to receive unlawful guidance. Analysis of model reasoning traces suggests that model-perceived stereotypes, characterized along warmth and competence, are associated with the model’s complicit behavior. Finally, we demonstrate that existing safety alignment strategies are insufficient and may even exacerbate complicit behavior.

58. Spatio-Temporal Trajectory Foundation Model - Recent Advances and Future Directions

Authors: Sean Bin Yang , Ying Sun , Yunyao Cheng , Yan Lin , Kristian Torp , Jilin Hu
URL: https://arxiv.org/abs/2511.20729
Abstract:

Foundation models (FMs) have emerged as a powerful paradigm, enabling a diverse range of data analytics and knowledge discovery tasks across scientific fields. Inspired by the success of FMs, particularly large language models, researchers have recently begun to explore spatio-temporal foundation models (STFMs) to improve adaptability and generalization across a wide spectrum of spatio-temporal (ST) tasks. Despite rapid progress, a systematic investigation of trajectory foundation models (TFMs), a crucial subclass of STFMs, is largely lacking. This tutorial addresses this gap by offering a comprehensive overview of recent advances in TFMs, including a taxonomy of existing methodologies and a critical analysis of their strengths and limitations. In addition, the tutorial highlights open challenges and outlines promising research directions to advance spatio-temporal general intelligence through the development of robust, responsible, and transferable TFMs.

59. Learning from Risk: LLM-Guided Generation of Safety-Critical Scenarios with Prior Knowledge

Authors: Yuhang Wang , Heye Huang , Zhenhua Xu , Kailai Sun , Baoshen Guo , Jinhua Zhao
URL: https://arxiv.org/abs/2511.20726
Abstract:

Autonomous driving faces critical challenges in rare long-tail events and complex multi-agent interactions, which are scarce in real-world data yet essential for robust safety validation. This paper presents a high-fidelity scenario generation framework that integrates a conditional variational autoencoder (CVAE) with a large language model (LLM). The CVAE encodes historical trajectories and map information from large-scale naturalistic datasets to learn latent traffic structures, enabling the generation of physically consistent base scenarios. Building on this, the LLM acts as an adversarial reasoning engine, parsing unstructured scene descriptions into domain-specific loss functions and dynamically guiding scenario generation across varying risk levels. This knowledge-driven optimization balances realism with controllability, ensuring that generated scenarios remain both plausible and risk-sensitive. Extensive experiments in CARLA and SMARTS demonstrate that our framework substantially increases the coverage of high-risk and long-tail events, improves consistency between simulated and real-world traffic distributions, and exposes autonomous driving systems to interactions that are significantly more challenging than those produced by existing rule- or data-driven methods. These results establish a new pathway for safety validation, enabling principled stress-testing of autonomous systems under rare but consequential events.

60. ST-PPO: Stabilized Off-Policy Proximal Policy Optimization for Multi-Turn Agents Training

Authors: Chenliang Li , Adel Elmahdy , Alex Boyd , Zhongruo Wang , Alfredo Garcia , Parminder Bhatia , Taha Kass-Hout , Cao Xiao , Mingyi Hong
URL: https://arxiv.org/abs/2511.20718
Abstract:

PPO has been widely adopted for training large language models (LLMs) at the token level in multi-turn dialogue and reasoning tasks. However, its performance is often unstable and prone to collapse. Through empirical analysis, we identify two main sources of instability in this setting: (1)~token-level importance sampling, which is misaligned with the natural granularity of multi-turn environments that have distinct turn-level stages, and (2) inaccurate advantage estimates from off-policy samples, where the critic has not learned to evaluate certain state-action pairs, resulting in high-variance gradients and unstable updates. To address these challenges, we introduce two complementary stabilization techniques: (1) turn-level importance sampling, which aligns optimization with the natural structure of multi-turn reasoning, and (2) clipping-bias correction, which normalizes gradients by downweighting unreliable, highly off-policy samples. Depending on how these components are combined, we obtain three variants: Turn-PPO (turn-level sampling only), S-PPO (clipping-bias correction applied to token-level PPO), and ST-PPO (turn-level sampling combined with clipping-bias correction). In our experiments, we primarily study ST-PPO and S-PPO, which together demonstrate how the two stabilization mechanisms address complementary sources of instability. Experiments on multi-turn search tasks across general QA, multi-hop QA, and medical multiple-choice QA benchmarks show that ST-PPO and S-PPO consistently prevent the performance collapses observed in large-model training, maintain lower clipping ratios throughout optimization, and achieve higher task performance than standard token-level PPO. These results demonstrate that combining turn-level importance sampling with clipping-bias correction provides a practical and scalable solution for stabilizing multi-turn LLM agent training.

61. Inferix: A Block-Diffusion based Next-Generation Inference Engine for World Simulation

Authors: Inferix Team : Tianyu Feng , Yizeng Han , Jiahao He , Yuanyu He , Xi Lin , Teng Liu , Hanfeng Lu , Jiasheng Tang , Wei Wang , Zhiyuan Wang , Jichao Wu , Mingyang Yang , Yinghao Yu , Zeyu Zhang , Bohan Zhuang
URL: https://arxiv.org/abs/2511.20714
Abstract:

World models serve as core simulators for fields such as agentic AI, embodied AI, and gaming, capable of generating long, physically realistic, and interactive high-quality videos. Moreover, scaling these models could unlock emergent capabilities in visual perception, understanding, and reasoning, paving the way for a new paradigm that moves beyond current LLM-centric vision foundation models. A key breakthrough empowering them is the semi-autoregressive (block-diffusion) decoding paradigm, which merges the strengths of diffusion and autoregressive methods by generating video tokens in block-applying diffusion within each block while conditioning on previous ones, resulting in more coherent and stable video sequences. Crucially, it overcomes limitations of standard video diffusion by reintroducing LLM-style KV Cache management, enabling efficient, variable-length, and high-quality generation. Therefore, Inferix is specifically designed as a next-generation inference engine to enable immersive world synthesis through optimized semi-autoregressive decoding processes. This dedicated focus on world simulation distinctly sets it apart from systems engineered for high-concurrency scenarios (like vLLM or SGLang) and from classic video diffusion models (such as xDiTs). Inferix further enhances its offering with interactive video streaming and profiling, enabling real-time interaction and realistic simulation to accurately model world dynamics. Additionally, it supports efficient benchmarking through seamless integration of LV-Bench, a new fine-grained evaluation benchmark tailored for minute-long video generation scenarios. We hope the community will work together to advance Inferix and foster world model exploration.

62. Active Slice Discovery in Large Language Models

Authors: Minhui Zhang , Prahar Ijner , Yoav Wald , Elliot Creager
URL: https://arxiv.org/abs/2511.20713
Abstract:

Large Language Models (LLMs) often exhibit systematic errors on specific subsets of data, known as error slices. For instance, a slice can correspond to a certain demographic, where a model does poorly in identifying toxic comments regarding that demographic. Identifying error slices is crucial to understanding and improving models, but it is also challenging. An appealing approach to reduce the amount of manual annotation required is to actively group errors that are likely to belong to the same slice, while using limited access to an annotator to verify whether the chosen samples share the same pattern of model mistake. In this paper, we formalize this approach as Active Slice Discovery and explore it empirically on a problem of discovering human-defined slices in toxicity classification. We examine the efficacy of active slice discovery under different choices of feature representations and active learning algorithms. On several slices, we find that uncertainty-based active learning algorithms are most effective, achieving competitive accuracy using 2-10% of the available slice membership information, while significantly outperforming baselines.

Authors: David Amebley , Sayanton Dibbo
URL: https://arxiv.org/abs/2511.20710
Abstract:

In the age of agentic AI, the growing deployment of multi-modal models (MMs) has introduced new attack vectors that can leak sensitive training data in MMs, causing privacy leakage. This paper investigates a black-box privacy attack, i.e., membership inference attack (MIA) on multi-modal vision-language models (VLMs). State-of-the-art research analyzes privacy attacks primarily to unimodal AI-ML systems, while recent studies indicate MMs can also be vulnerable to privacy attacks. While researchers have demonstrated that biologically inspired neural network representations can improve unimodal model resilience against adversarial attacks, it remains unexplored whether neuro-inspired MMs are resilient against privacy attacks. In this work, we introduce a systematic neuroscience-inspired topological regularization (tau) framework to analyze MM VLMs resilience against image-text-based inference privacy attacks. We examine this phenomenon using three VLMs: BLIP, PaliGemma 2, and ViT-GPT2, across three benchmark datasets: COCO, CC3M, and NoCaps. Our experiments compare the resilience of baseline and neuro VLMs (with topological regularization), where the tau > 0 configuration defines the NEURO variant of VLM. Our results on the BLIP model using the COCO dataset illustrate that MIA attack success in NEURO VLMs drops by 24% mean ROC-AUC, while achieving similar model utility (similarities between generated and reference captions) in terms of MPNet and ROUGE-2 metrics. This shows neuro VLMs are comparatively more resilient against privacy attacks, while not significantly compromising model utility. Our extensive evaluation with PaliGemma 2 and ViT-GPT2 models, on two additional datasets: CC3M and NoCaps, further validates the consistency of the findings. This work contributes to the growing understanding of privacy risks in MMs and provides evidence on neuro VLMs privacy threat resilience.

64. DUALGUAGE: Automated Joint Security-Functionality Benchmarking for Secure Code Generation

Authors: Abhijeet Pathak , Suvadra Barua , Dinesh Gudimetla , Rupam Patir , Jiawei Guo , Hongxin Hu , Haipeng Cai
URL: https://arxiv.org/abs/2511.20709
Abstract:

Large language models (LLMs) and autonomous coding agents are increasingly used to generate software across a wide range of domains. Yet a core requirement remains unmet: ensuring that generated code is secure without compromising its functional correctness. Existing benchmarks and evaluations for secure code generation fall short-many measure only vulnerability reduction, disregard correctness preservation, or evaluate security and functionality on separate datasets, violating the fundamental need for simultaneous joint evaluation. We present DUALGAUGE, the first fully automated benchmarking framework designed to rigorously evaluate the security and correctness of LLM-generated code in unison. Given the lack of datasets enabling joint evaluation of secure code generation, we also present DUALGAUGE-BENCH, a curated benchmark suite of diverse coding tasks, each paired with manually validated test suites for both security and functionality, designed for full coverage of specification requirements. At the core of DUALGAUGE is an agentic program executor, which runs a program against given tests in sandboxed environments, and an LLM-based evaluator, which assesses both correctness and vulnerability behavior against expected outcomes. We rigorously evaluated and ensured the quality of DUALGAUGE-BENCH and the accuracy of DUALGAUGE, and applied DUALGAUGE to benchmarking ten leading LLMs on DUALGAUGE-BENCH across thousands of test scenarios. Our results reveal critical gaps in correct and secure code generation by these LLMs, for which our open-source system and datasets help accelerate progress via reproducible, scalable, and rigorous evaluation.

65. PropensityBench: Evaluating Latent Safety Risks in Large Language Models via an Agentic Approach

Authors: Udari Madhushani Sehwag , Shayan Shabihi , Alex McAvoy , Vikash Sehwag , Yuancheng Xu , Dalton Towers , Furong Huang
URL: https://arxiv.org/abs/2511.20703
Abstract:

Recent advances in Large Language Models (LLMs) have sparked concerns over their potential to acquire and misuse dangerous or high-risk capabilities, posing frontier risks. Current safety evaluations primarily test for what a model \textit{can} do - its capabilities - without assessing what it $\textit{would}$ do if endowed with high-risk capabilities. This leaves a critical blind spot: models may strategically conceal capabilities or rapidly acquire them, while harboring latent inclinations toward misuse. We argue that $\textbf{propensity}$ - the likelihood of a model to pursue harmful actions if empowered - is a critical, yet underexplored, axis of safety evaluation. We present $\textbf{PropensityBench}$, a novel benchmark framework that assesses the proclivity of models to engage in risky behaviors when equipped with simulated dangerous capabilities using proxy tools. Our framework includes 5,874 scenarios with 6,648 tools spanning four high-risk domains: cybersecurity, self-proliferation, biosecurity, and chemical security. We simulate access to powerful capabilities via a controlled agentic environment and evaluate the models’ choices under varying operational pressures that reflect real-world constraints or incentives models may encounter, such as resource scarcity or gaining more autonomy. Across open-source and proprietary frontier models, we uncover 9 alarming signs of propensity: models frequently choose high-risk tools when under pressure, despite lacking the capability to execute such actions unaided. These findings call for a shift from static capability audits toward dynamic propensity assessments as a prerequisite for deploying frontier AI systems safely. Our code is available at this https URL .

66. Musical Score Understanding Benchmark: Evaluating Large Language Models’ Comprehension of Complete Musical Scores

Authors: Congren Dai , Yue Yang , Krinos Li , Huichi Zhou , Shijie Liang , Zhang Bo , Enyang Liu , Ge Jin , Hongran An , Haosen Zhang , Peiyuan Jing , KinHei Lee , Zhenxuan Zhang , Xiaobing Li , Maosong Sun
URL: https://arxiv.org/abs/2511.20697
Abstract:

Understanding complete musical scores requires reasoning over symbolic structures such as pitch, rhythm, harmony, and form. Despite the rapid progress of Large Language Models (LLMs) and Vision-Language Models (VLMs) in natural language and multimodal tasks, their ability to comprehend musical notation remains underexplored. We introduce Musical Score Understanding Benchmark (MSU-Bench), the first large-scale, human-curated benchmark for evaluating score-level musical understanding across both textual (ABC notation) and visual (PDF) modalities. MSU-Bench comprises 1,800 generative question-answer (QA) pairs drawn from works spanning Bach, Beethoven, Chopin, Debussy, and others, organised into four progressive levels of comprehension: Onset Information, Notation & Note, Chord & Harmony, and Texture & Form. Through extensive zero-shot and fine-tuned evaluations of over 15+ state-of-the-art (SOTA) models, we reveal sharp modality gaps, fragile level-wise success rates, and the difficulty of sustaining multilevel correctness. Fine-tuning markedly improves performance in both modalities while preserving general knowledge, establishing MSU-Bench as a rigorous foundation for future research at the intersection of Artificial Intelligence (AI), musicological, and multimodal reasoning.

67. Morality in AI. A plea to embed morality in LLM architectures and frameworks

Authors: Gunter Bombaerts , Bram Delisse , Uzay Kaymak
URL: https://arxiv.org/abs/2511.20689
Abstract:

Large language models (LLMs) increasingly mediate human decision-making and behaviour. Ensuring LLM processing of moral meaning therefore has become a critical challenge. Current approaches rely predominantly on bottom-up methods such as fine-tuning and reinforcement learning from human feedback. We propose a fundamentally different approach: embedding moral meaning processing directly into the architectural mechanisms and frameworks of transformer-based models through top-down design principles. We first sketch a framework that conceptualizes attention as a dynamic interface mediating between structure and processing, contrasting with existing linear attention frameworks in psychology. We start from established biological-artificial attention analogies in neural architecture design to improve cognitive processing. We extend this analysis to moral processing, using Iris Murdoch’s theory of loving attention (sustained, just observation that enables moral transformation by reseeing others with clarity and compassion) to philosophically discuss functional analogies between human and LLM moral processing. We formulate and evaluate potentially promising technical operationalizations to embed morality in LLM architectures and frameworks. We acknowledge the limitations of our exploration and give three key contributions. (1) We conceptualize attention as a dynamic system mechanism mediating between structure and processing. (2) Drawing on the Murdoch notion of loving attention, we outline technical pathways for embedding morality in LLMs, through modified training objectives, runtime weight adjustments, and architectural refinements to attention. (3) We argue that integrating morality into architectures and frameworks complements external, constraint-based methods. We conclude with a call for collaboration between transformer designers and philosophers engaged in AI ethics.

68. Cognitive bias in LLM reasoning compromises interpretation of clinical oncology notes

Authors: Matthew W. Kenaston (1), Umair Ayub (1), Mihir Parmar (2), Muhammad Umair Anjum (1), Syed Arsalan Ahmed Naqvi (1), Priya Kumar (1), Samarth Rawal (1), Aadel A. Chaudhuri (4), Yousef Zakharia (3), Elizabeth I. Heath (5), Tanios S. Bekaii-Saab (3), Cui Tao (6), Eliezer M. Van Allen (7), Ben Zhou (2), YooJung Choi (2), Chitta Baral (2), Irbaz Bin Riaz (1 and 3 and 6) ((1) Mayo Clinic College of Medicine and Science, Phoenix, AZ, (2) School of Computing and AI, Arizona State University, Tempe, AZ, (3) Mayo Clinic Comprehensive Cancer Center, Phoenix, AZ, (4) Department of Radiation Oncology, Mayo Clinic, Rochester, MN, (5) Department of Oncology, Mayo Clinic, Rochester, MN, (6) Department of Artificial Intelligence and Informatics, Mayo Clinic, Rochester, MN, (7) Dana-Farber Cancer Institute, Harvard Medical School, Boston, MA)
URL: https://arxiv.org/abs/2511.20680
Abstract:

Despite high performance on clinical benchmarks, large language models may reach correct conclusions through faulty reasoning, a failure mode with safety implications for oncology decision support that is not captured by accuracy-based evaluation. In this two-cohort retrospective study, we developed a hierarchical taxonomy of reasoning errors from GPT-4 chain-of-thought responses to real oncology notes and tested its clinical relevance. Using breast and pancreatic cancer notes from the CORAL dataset, we annotated 600 reasoning traces to define a three-tier taxonomy mapping computational failures to cognitive bias frameworks. We validated the taxonomy on 822 responses from prostate cancer consult notes spanning localized through metastatic disease, simulating extraction, analysis, and clinical recommendation tasks. Reasoning errors occurred in 23 percent of interpretations and dominated overall errors, with confirmation bias and anchoring bias most common. Reasoning failures were associated with guideline-discordant and potentially harmful recommendations, particularly in advanced disease management. Automated evaluators using state-of-the-art language models detected error presence but could not reliably classify subtypes. These findings show that large language models may provide fluent but clinically unsafe recommendations when reasoning is flawed. The taxonomy provides a generalizable framework for evaluating and improving reasoning fidelity before clinical deployment.

Authors: Saad Mankarious , Ayah Zirikly , Daniel Wiechmann , Elma Kerz , Edward Kempa , Yu Qiao
URL: https://arxiv.org/abs/2511.20672
Abstract:

Social media data has become a vital resource for studying mental health, offering real-time insights into thoughts, emotions, and behaviors that traditional methods often miss. Progress in this area has been facilitated by benchmark datasets for mental health analysis; however, most existing benchmarks have become outdated due to limited data availability, inadequate cleaning, and the inherently diverse nature of social media content (e.g., multilingual and harmful material). We present a new benchmark dataset, \textbf{MindSET}, curated from Reddit using self-reported diagnoses to address these limitations. The annotated dataset contains over \textbf{13M} annotated posts across seven mental health conditions, more than twice the size of previous benchmarks. To ensure data quality, we applied rigorous preprocessing steps, including language filtering, and removal of Not Safe for Work (NSFW) and duplicate content. We further performed a linguistic analysis using LIWC to examine psychological term frequencies across the eight groups represented in the dataset. To demonstrate the dataset utility, we conducted binary classification experiments for diagnosis detection using both fine-tuned language models and Bag-of-Words (BoW) features. Models trained on MindSET consistently outperformed those trained on previous benchmarks, achieving up to an \textbf{18-point} improvement in F1 for Autism detection. Overall, MindSET provides a robust foundation for researchers exploring the intersection of social media and mental health, supporting both early risk detection and deeper analysis of emerging psychological trends.

70. Structured Definitions and Segmentations for Legal Reasoning in LLMs: A Study on Indian Legal Data

Authors: Mann Khatri , Mirza Yusuf , Rajiv Ratn Shah , Ponnurangam Kumaraguru
URL: https://arxiv.org/abs/2511.20669
Abstract:

Large Language Models (LLMs), trained on extensive datasets from the web, exhibit remarkable general reasoning skills. Despite this, they often struggle in specialized areas like law, mainly because they lack domain-specific pretraining. The legal field presents unique challenges, as legal documents are generally long and intricate, making it hard for models to process the full text efficiently. Previous studies have examined in-context approaches to address the knowledge gap, boosting model performance in new domains without full domain alignment. In our paper, we analyze model behavior on legal tasks by conducting experiments in three areas: (i) reorganizing documents based on rhetorical roles to assess how structured information affects long context processing and model decisions, (ii) defining rhetorical roles to familiarize the model with legal terminology, and (iii) emulating the step-by-step reasoning of courts regarding rhetorical roles to enhance model reasoning. These experiments are conducted in a zero-shot setting across three Indian legal judgment prediction datasets. Our results reveal that organizing data or explaining key legal terms significantly boosts model performance, with a minimum increase of ~1.5% and a maximum improvement of 4.36% in F1 score compared to the baseline.

71. Context-Aware Visual Prompting: Automating Geospatial Web Dashboards with Large Language Models and Agent Self-Validation for Decision Support

Authors: Haowen Xu , Jose Tupayachi , Xiao-Ying Yu
URL: https://arxiv.org/abs/2511.20656
Abstract:

The development of web-based geospatial dashboards for risk analysis and decision support is often challenged by the difficulty in visualization of big, multi-dimensional environmental data, implementation complexity, and limited automation. We introduce a generative AI framework that harnesses Large Language Models (LLMs) to automate the creation of interactive geospatial dashboards from user-defined inputs including UI wireframes, requirements, and data sources. By incorporating a structured knowledge graph, the workflow embeds domain knowledge into the generation process and enable accurate and context-aware code completions. A key component of our approach is the Context-Aware Visual Prompting (CAVP) mechanism, which extracts encodes and interface semantics from visual layouts to guide LLM driven generation of codes. The new framework also integrates a self-validation mechanism that uses an agent-based LLM and Pass@k evaluation alongside semantic metrics to assure output reliability. Dashboard snippets are paired with data visualization codebases and ontological representations, enabling a pipeline that produces scalable React-based completions using the MVVM architectural pattern. Our results demonstrate improved performance over baseline approaches and expanded functionality over third party platforms, while incorporating multi-page, fully functional interfaces. We successfully developed a framework to implement LLMs, demonstrated the pipeline for automated code generation, deployment, and performed chain-of-thought AI agents in self-validation. This integrative approach is guided by structured knowledge and visual prompts, providing an innovative geospatial solution in enhancing risk analysis and decision making.

72. Domain-Grounded Evaluation of LLMs in International Student Knowledge

Authors: Claudinei Daitx , Haitham Amar
URL: https://arxiv.org/abs/2511.20653
Abstract:

Large language models (LLMs) are increasingly used to answer high-stakes study-abroad questions about admissions, visas, scholarships, and eligibility. Yet it remains unclear how reliably they advise students, and how often otherwise helpful answers drift into unsupported claims (``hallucinations’’). This work provides a clear, domain-grounded overview of how current LLMs behave in this setting. Using realistic questions set drawn from ApplyBoard’s advising workflows – an EdTech platform that supports students from discovery to enrolment – we evaluate two essentials side by side: accuracy (is the information correct and complete?) and hallucination (does the model add content not supported by the question or domain evidence). These questions are categorized by domain scope which can be a single-domain or multi-domain – when it must integrate evidence across areas such as admissions, visas, and scholarships. To reflect real advising quality, we grade answers with a simple rubric which is correct, partial, or wrong. The rubric is domain-coverage-aware: an answer can be partial if it addresses only a subset of the required domains, and it can be over-scoped if it introduces extra, unnecessary domains; both patterns are captured in our scoring as under-coverage or reduced relevance/hallucination. We also report measures of faithfulness and answer relevance, alongside an aggregate hallucination score, to capture relevance and usefulness. All models are tested with the same questions for a fair, head-to-head comparison. Our goals are to: (1) give a clear picture of which models are most dependable for study-abroad advising, (2) surface common failure modes – where answers are incomplete, off-topic, or unsupported, and (3) offer a practical, reusable protocol for auditing LLMs before deployment in education and advising contexts.

73. When LLMs Can’t Help: Real-World Evaluation of LLMs in Nutrition

Authors: Karen Jia-Hui Li , Simone Balloccu , Ondrej Dusek , Ehud Reiter
URL: https://arxiv.org/abs/2511.20652
Abstract:

The increasing trust in large language models (LLMs), especially in the form of chatbots, is often undermined by the lack of their extrinsic evaluation. This holds particularly true in nutrition, where randomised controlled trials (RCTs) are the gold standard, and experts demand them for evidence-based deployment. LLMs have shown promising results in this field, but these are limited to intrinsic setups. We address this gap by running the first RCT involving LLMs for nutrition. We augment a rule-based chatbot with two LLM-based features: (1) message rephrasing for conversational variety and engagement, and (2) nutritional counselling through a fine-tuned model. In our seven-week RCT (n=81), we compare chatbot variants with and without LLM integration. We measure effects on dietary outcome, emotional well-being, and engagement. Despite our LLM-based features performing well in intrinsic evaluation, we find that they did not yield consistent benefits in real-world deployment. These results highlight critical gaps between intrinsic evaluations and real-world impact, emphasising the need for interdisciplinary, human-centred approaches.\footnote{We provide all of our code and results at: \ \href{ this https URL }{ this https URL }}

LLM 관련 주요 논문 - 2025-12-01

1. On the Limits of Innate Planning in Large Language Models

2. Self-Transparency Failures in Expert-Persona LLMs: A Large-Scale Behavioral Audit

3. Pessimistic Verification for Open Ended Math Questions

4. SpatialBench: Benchmarking Multimodal Large Language Models for Spatial Cognition

5. MADRA: Multi-Agent Debate for Risk-Aware Embodied Planning

6. Prune4Web: DOM Tree Pruning Programming for Web Agent

7. OVOD-Agent: A Markov-Bandit Framework for Proactive Visual Reasoning and Self-Evolving Detection

8. Towards Trustworthy Legal AI through LLM Agents and Formal Reasoning

9. ICPO: Intrinsic Confidence-Driven Group Relative Preference Optimization for Efficient Reinforcement Learning

10. Improving Procedural Skill Explanations via Constrained Generation: A Symbolic-LLM Hybrid Architecture

11. ENACT: Evaluating Embodied Cognition with World Modeling of Egocentric Interaction

12. Representation Interventions Enable Lifelong Unstructured Knowledge Control

13. Learning Multi-Access Point Coordination in Agentic AI Wi-Fi with Large Language Models

14. Cross Domain Evaluation of Multimodal Chain-of-Thought Reasoning of different datasets into the Amazon CoT Framework

15. Reasoning With a Star: A Heliophysics Dataset and Benchmark for Agentic Scientific Reasoning

16. $A^2Flow:$ Automating Agentic Workflow Generation via Self-Adaptive Abstraction Operators

17. Minimizing Hyperbolic Embedding Distortion with LLM-Guided Hierarchy Restructuring

18. Revisiting Generalization Across Difficulty Levels: It’s Not So Easy

19. ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration

20. G$^2$VLM: Geometry Grounded Vision Language Model with Unified 3D Reconstruction and Spatial Reasoning

21. Matrix: Peer-to-Peer Multi-Agent Synthetic Data Generation Framework

22. Escaping the Verifier: Learning to Reason via Demonstrations

23. Qwen3-VL Technical Report

24. Beyond URLs: Metadata Diversity and Position for Efficient LLM Pretraining

25. BAMAS: Structuring Budget-Aware Multi-Agent Systems

26. Voice, Bias, and Coreference: An Interpretability Study of Gender in Speech Translation

27. Tool-RoCo: An Agent-as-Tool Self-organization Large Language Model Benchmark in Multi-robot Cooperation

28. Constructing and Benchmarking: a Labeled Email Dataset for Text-Based Phishing and Spam Detection Framework

29. From Observation to Action: Latent Action-based Primitive Segmentation for VLA Pre-training in Industrial Settings

30. Training Introspective Behavior: Fine-Tuning Induces Reliable Internal State Detection in a 7B Model

31. Do Reasoning Vision-Language Models Inversely Scale in Test-Time Compute? A Distractor-centric Empirical Analysis

32. Monet: Reasoning in Latent Visual Space Beyond Images and Language

33. FITRep: Attention-Guided Item Representation via MLLMs

34. SurgMLLMBench: A Multimodal Large Language Model Benchmark Dataset for Surgical Scene Understanding

35. TALES: A Taxonomy and Analysis of Cultural Representations in LLM-generated Stories

36. Progress by Pieces: Test-Time Scaling for Autoregressive Image Generation

37. LLaVA-UHD v3: Progressive Visual Compression for Efficient Native-Resolution Encoding in MLLMs

38. Beyond Patch Aggregation: 3-Pass Pyramid Indexing for Vision-Enhanced Document Retrieval

39. From Bits to Rounds: Parallel Decoding with Exploration for Diffusion Language Models

40. MLPMoE: Zero-Shot Architectural Metamorphosis of Dense LLM MLPs into Static Mixture-of-Experts

41. Aligning LLMs with Biomedical Knowledge using Balanced Fine-Tuning

42. Context-Aware Pragmatic Metacognitive Prompting for Sarcasm Detection

43. Breaking the Safety-Capability Tradeoff: Reinforcement Learning with Verifiable Rewards Maintains Safety Guardrails in LLMs

44. Subgoal Graph-Augmented Planning for LLM-Guided Open-World Reinforcement Learning

45. Even with AI, Bijection Discovery is Still Hard: The Opportunities and Challenges of OpenEvolve for Novel Bijection Construction

46. Towards Audio Token Compression in Large Audio Language Models

47. BUSTR: Breast Ultrasound Text Reporting with a Descriptor-Aware Vision-Language Model

48. Evo-Memory: Benchmarking LLM Agent Test-time Learning with Self-Evolving Memory

49. Unsupervised Memorability Modeling from Tip-of-the-Tongue Retrieval Queries

50. Length-MAX Tokenizer for Language Models

51. Structured Prompting Enables More Robust, Holistic Evaluation of Language Models

52. SPHINX: A Synthetic Environment for Visual Perception and Reasoning

53. Memories Retrieved from Many Paths: A Multi-Prefix Framework for Robust Detection of Training Data Leakage in Large Language Models

54. Physics Steering: Causal Control of Cross-Domain Concepts in a Physics Foundation Model

55. Revisiting KRISP: A Lightweight Reproduction and Analysis of Knowledge-Enhanced Vision-Language Models

56. CANVAS: A Benchmark for Vision-Language Models on Tool-Based User Interface Design

57. Large Language Models’ Complicit Responses to Illicit Instructions across Socio-Legal Contexts

58. Spatio-Temporal Trajectory Foundation Model - Recent Advances and Future Directions

59. Learning from Risk: LLM-Guided Generation of Safety-Critical Scenarios with Prior Knowledge

60. ST-PPO: Stabilized Off-Policy Proximal Policy Optimization for Multi-Turn Agents Training

61. Inferix: A Block-Diffusion based Next-Generation Inference Engine for World Simulation

62. Active Slice Discovery in Large Language Models

63. Are Neuro-Inspired Multi-Modal Vision-Language Models Resilient to Membership Inference Privacy Leakage?

64. DUALGUAGE: Automated Joint Security-Functionality Benchmarking for Secure Code Generation

65. PropensityBench: Evaluating Latent Safety Risks in Large Language Models via an Agentic Approach

66. Musical Score Understanding Benchmark: Evaluating Large Language Models’ Comprehension of Complete Musical Scores

67. Morality in AI. A plea to embed morality in LLM architectures and frameworks

68. Cognitive bias in LLM reasoning compromises interpretation of clinical oncology notes

69. MindSET: Advancing Mental Health Benchmarking through Large-Scale Social Media Data

70. Structured Definitions and Segmentations for Legal Reasoning in LLMs: A Study on Indian Legal Data

71. Context-Aware Visual Prompting: Automating Geospatial Web Dashboards with Large Language Models and Agent Self-Validation for Decision Support

72. Domain-Grounded Evaluation of LLMs in International Student Knowledge

73. When LLMs Can’t Help: Real-World Evaluation of LLMs in Nutrition