Searcharxiv⌕ Search

arXiv subjects

Sharon Li

Publications and source records attributed to Sharon Li.

At least 37 records · Page 2Linked to original sources
arXiv2026

DECOR: Auditing LLM Deception via Information Manipulation Theory

Large language models can deceive by subtly manipulating truthful information -- omitting key facts, shifting focus, or obscuring meaning -- making such behavior difficult to detect. Existing black-box methods rely on coarse-grained judgments, offering limited interpretability and failing to pinpoint which facts were distorted and how. We introduce DECOR, a multi-agent framework grounded in Information Manipulation Theory for fine-grained auditing of strategic deception in LLM responses. DECOR decomposes input contexts into atomic informational units and scores each unit against the response across four dimensions of manipulation, producing interpretable manipulation profiles that are aggregated into a global deception index. We comprehensively evaluate DECOR on both single-turn and multi-turn deception detection benchmarks spanning real-world domains, and show that DECOR achieves state-of-the-art performance on both, outperforming competitive baselines. The framework generalizes across 15 frontier models, and ablation studies confirm the contribution of each key design component. Our findings demonstrate that fine-grained, theory-grounded auditing of information manipulation offers an effective and interpretable path for LLM deception detection.

cs.CL↗
arXiv2026

Auditing Agent Harness Safety

LLM agents increasingly run inside execution harnesses that dispatch tools, allocate resources, and route messages between specialized components. However, a harness can return a correct, benign answer over a trajectory that accesses unauthorized resources or leaks context to the wrong agent. Output-level evaluation cannot see these failures, yet most safety benchmarks score only final outputs or terminal states, even though many violations occur mid-trajectory rather than at termination. The central question is whether the harness respects user intent, permission boundaries, and information-flow constraints throughout execution. To address this gap, we propose HarnessAudit, a framework that audits full execution trajectories across boundary compliance, execution fidelity, and system stability, with a focus on multi-agent harnesses where these risks are most pronounced. We further introduce HarnessAudit-Bench, a benchmark of 210 tasks across eight real-world domains, instantiated in both single-agent and multi-agent configurations with embedded safety constraints. Evaluating ten harness configurations across frontier models and three multi-agent frameworks, we find that: (i) task completion is misaligned with safe execution, and violations accumulate with trajectory length; (ii) safety risks vary across domains, task types, and agent roles; (iii) most violations concentrate in resource access and inter-agent information transfer; and (iv) multi-agent collaboration expands the safety risk surface, while harness design sets the upper bound of safe deployment.

cs.CL↗
arXiv2026

How Do Transformers Learn to Associate Tokens: Gradient Leading Terms Bring Mechanistic Interpretability

Semantic associations such as the link between "bird" and "flew" are foundational for language modeling as they enable models to go beyond memorization and instead generalize and generate coherent text. Understanding how these associations are learned and represented in language models is essential for connecting deep learning with linguistic theory and developing a mechanistic foundation for large language models. In this work, we analyze how these associations emerge from natural language data in attention-based language models through the lens of training dynamics. By leveraging a leading-term approximation of the gradients, we develop closed-form expressions for the weights at early stages of training that explain how semantic associations first take shape. Through our analysis, we reveal that each set of weights of the transformer has closed-form expressions as simple compositions of three basis functions (bigram, token-interchangeability, and context mappings), reflecting the statistics of the text corpus and uncovering how each component of the transformer captures semantic associations based on these compositions. Experiments on real-world LLMs demonstrate that our theoretical weight characterizations closely match the learned weights, and qualitative analyses further show how our theorem shines light on interpreting the learned associations in transformers.

cs.CL↗
arXiv2026

Detecting Distillation Data from Reasoning Models

Reasoning distillation has emerged as a prevailing paradigm for transferring reasoning capabilities from large reasoning models to small language models. Yet, reasoning distillation risks data contamination: benchmark data may inadvertently be included in the distillation data, thereby inflating model performance metrics. In this work, we formally define the distillation data detection task, which determines whether a given question is included in the model's distillation data. The unique challenge of this task lies in the partial availability of distillation data. To address this, we propose Token Probability Deviation (TPD), a detection method that leverages the probability patterns of output tokens generated by the model instead of input tokens. Our method is motivated by the observation that seen questions tend to elicit more near-deterministic tokens generated by the models than unseen ones. Our TPD score is thus designed to quantify the token-level deviation of generated tokens from a high-confidence reference probability. Consequently, seen questions can yield substantially lower TPD scores than unseen ones, enabling strong detection performance. Extensive experiments demonstrate the effectiveness of our approach, improving detection AUC by up to 31% on distillation datasets.

cs.CL↗
arXiv2026

VAUQ: Vision-Aware Uncertainty Quantification for LVLM Self-Evaluation

Large Vision-Language Models (LVLMs) frequently hallucinate, limiting their safe deployment in real-world applications. Existing LLM self-evaluation methods rely on a model's ability to estimate the correctness of its own outputs, which can improve deployment reliability; however, they depend heavily on language priors and are therefore ill-suited for evaluating vision-conditioned predictions. We propose VAUQ, a vision-aware uncertainty quantification framework for LVLM self-evaluation that explicitly measures how strongly a model's output depends on visual evidence. VAUQ introduces the Image-Information Score (IS), which captures the reduction in predictive uncertainty attributable to visual input, and an unsupervised core-region masking strategy that amplifies the influence of salient regions. Combining predictive entropy with this core-masked IS yields a training-free scoring function that reliably reflects answer correctness. Comprehensive experiments show that VAUQ consistently outperforms existing self-evaluation methods across multiple datasets.

cs.CV↗
arXiv2026

Uncertainty Quantification in LLM Agents: Foundations, Emerging Challenges, and Opportunities

Uncertainty quantification (UQ) for large language models (LLMs) is a key building block for safety guardrails of daily LLM applications. Yet, even as LLM agents are increasingly deployed in highly complex tasks, most UQ research still centers on single-turn question-answering. We argue that UQ research must shift to realistic settings with interactive agents, and that a new principled framework for agent UQ is needed. This paper presents three pillars to build a solid ground for future agent UQ research: (1. Foundations) We present the first general formulation of agent UQ that subsumes broad classes of existing UQ setups; (2. Challenges) We identify four technical challenges specifically tied to agentic setups -- selection of uncertainty estimator, uncertainty of heterogeneous entities, modeling uncertainty dynamics in interactive systems, and lack of fine-grained benchmarks -- with numerical analysis on a real-world agent benchmark, $τ^2$-bench; (3. Future Directions) We conclude with noting on the practical implications of agent UQ and remaining open problems as forward-looking discussion for future explorations.

cs.AI↗
arXiv2026

GeoArena: Evaluating Open-World Geographic Reasoning in Large Vision-Language Models

Geographic reasoning is a fundamental cognitive capability that requires models to infer plausible locations by synthesizing visual evidence with spatial world knowledge. Despite recent advances in large vision-language models (LVLMs), existing evaluation paradigms remain largely outcome-centric, relying on static datasets and predefined labels that are conceptually misaligned with open-world geographic inference. Such outcome-centric evaluations often focus exclusively on label matching, leaving the underlying linguistic reasoning chains as unexamined black boxes. In this work, we introduce GeoArena, a dynamic, human-preference-based evaluation framework for benchmarking open-world geographic reasoning. GeoArena reframes evaluation as a pairwise reasoning alignment task on in-the-wild images, where human judges compare model-generated explanations based on reasoning quality, evidence synthesis, and plausibility. We deploy GeoArena as a public platform and benchmark 17 frontier LVLMs using thousands of human judgments, which complements existing benchmarks and supports the development of geographically grounded, human-aligned AI systems. We further provide detailed analyses of model behavior, including reliability of human preferences and factors influencing judgments of geographic reasoning quality.

cs.CV↗
arXiv2026

How Retrieved Context Shapes Internal Representations in RAG

Retrieval-augmented generation (RAG) enhances large language models (LLMs) by conditioning generation on retrieved external documents, but the effect of retrieved context is often non-trivial. In realistic retrieval settings, the retrieved document set often contains a mixture of documents that vary in relevance and usefulness. While prior work has largely examined these phenomena through output behavior, little is known about how retrieved context shapes the internal representations that mediate information integration in RAG. In this work, we study RAG through the lens of latent representations. We systematically analyze how different types of retrieved documents affect the hidden states of LLMs, and how these internal representation shifts relate to downstream generation behavior. Across four question-answering datasets and three LLMs, we analyze internal representations under controlled single- and multi-document settings. Our results reveal how context relevancy and layer-wise processing influence internal representations, providing explanations of LLMs' output behaviors and insights for RAG system design.

cs.CL↗
arXiv2026

Why Multimodal In-Context Learning Lags Behind? Unveiling the Inner Mechanisms and Bottlenecks

In-context learning (ICL) enables models to adapt to new tasks via inference-time demonstrations. Despite its success in large language models, the extension of ICL to multimodal settings remains poorly understood in terms of its internal mechanisms and how it differs from text-only ICL. In this work, we conduct a systematic analysis of ICL in multimodal large language models. Using identical task formulations across modalities, we show that multimodal ICL performs comparably to text-only ICL in zero-shot settings but degrades significantly under few-shot demonstrations. To understand this gap, we decompose multimodal ICL into task mapping construction and task mapping transfer, and analyze how models establish cross-modal task mappings, and transfer them to query samples across layers. Our analysis reveals that current models lack reasoning-level alignment between visual and textual representations, and fail to reliably transfer learned task mappings to queries. Guided by these findings, we further propose a simple inference-stage enhancement method that reinforces task mapping transfer. Our results provide new insights into the mechanisms and limitations of multimodal ICL and suggest directions for more effective multimodal adaptation. Our code is available \href{https://github.com/deeplearning-wisc/Multimocal-ICL-Analysis-Framework-MGI}{here}.

cs.CV↗
arXiv2026

Unsupervised Domain Adaptation for Binary Classification with an Unobservable Source Subpopulation

We study an unsupervised domain adaptation problem where the source domain consists of subpopulations defined by the binary label $Y$ and a binary background (or environment) $A$. We focus on a challenging setting in which one such subpopulation in the source domain is unobservable. Naively ignoring this unobserved group can result in biased estimates and degraded predictive performance. Despite this structured missingness, we show that the prediction in the target domain can still be recovered. Specifically, we rigorously derive both background-specific and overall prediction models for the target domain. For practical implementation, we propose the distribution matching method to estimate the subpopulation proportions. We provide theoretical guarantees for the asymptotic behavior of our estimator, and establish an upper bound on the prediction error. Experiments on both synthetic and real-world datasets show that our method outperforms the naive benchmark that does not account for this unobservable source subpopulation.

stat.ML↗
arXiv2026

When Identity Skews Debate: Anonymization for Bias-Reduced Multi-Agent Reasoning

Multi-agent debate (MAD) aims to improve large language model (LLM) reasoning by letting multiple agents exchange answers and then aggregate their opinions. Yet recent studies reveal that agents are not neutral: they are prone to identity-driven sycophancy and self-bias, uncritically adopting a peer's view or stubbornly adhering to their own prior output, undermining the reliability of debate. In this work, we present the first principled framework that joins sycophancy and self-bias to mitigate and quantify identity bias in MAD. First, we formalize the debate dynamics as an identity-weighted Bayesian update process. Second, we propose response anonymization: by removing identity markers from prompts, agents cannot distinguish "self" from "peer", which forces equal weights on agent identity, thereby reducing bias and improving trustworthiness. Third, we define the Identity Bias Coefficient (IBC), a principled bias metric that measures an agent's tendency to follow its peer versus itself. Empirical studies across multiple models and benchmarks confirm that identity bias is widespread, with sycophancy far more common than self-bias. Our findings highlight the need to ensure that MAD systems reason based on content rather than identity. Code is released in https://github.com/deeplearning-wisc/MAD-identity-bias.

cs.AI↗
arXiv2026

ModeX: Evaluator-Free Best-of-N Selection for Open-Ended Generation

Selecting a single high-quality output from multiple stochastic generations remains a fundamental challenge for large language models (LLMs), particularly in open-ended tasks where no canonical answer exists. While Best-of-N and self-consistency methods show that aggregating multiple generations can improve performance, existing approaches typically rely on external evaluators, reward models, or exact string-match voting, limiting their applicability and efficiency. We propose Mode Extraction (ModeX), an evaluator-free Best-of-N selection framework that generalizes majority voting to open-ended text generation by identifying the modal output representing the dominant semantic consensus among generated texts. ModeX constructs a similarity graph over candidate generations and recursively applies spectral clustering to select a representative centroid, without requiring additional inference or auxiliary models. We further instantiate this selection principle as ModeX-Lite, an improved version of ModeX with early pruning for efficiency. Across open-ended tasks -- including text summarization, code generation, and mathematical reasoning -- our approaches consistently outperform standard single- and multi-path baselines, providing a computationally efficient solution for robust open-ended text generation. Code is released in https://github.com/deeplearning-wisc/ModeX.

cs.CL↗
arXiv2026

Can LLMs Learn to Reason Robustly under Noisy Supervision?

Reinforcement Learning with Verifiable Rewards (RLVR) effectively trains reasoning models that rely on abundant perfect labels, but its vulnerability to unavoidable noisy labels due to expert scarcity remains critically underexplored. In this work, we take the first step toward a systematic analysis of noisy label mechanisms in RLVR. In contrast to supervised classification, most RLVR algorithms incorporate a rollout-based condition: a label's influence on training is contingent on whether the current policy can generate rollouts that realize it, a property that naturally extends to noisy labels. Based on this observation, we distinguish two types of noise: inactive noisy labels, which reduce data efficiency, and active noisy labels, which are reinforced and risk skewing the model toward incorrect distributions. From experiments on training with noisy samples, we identify an Early Correctness Coherence phenomenon: although noisy samples begin to lag behind in later stages, accuracy on both clean and noisy samples increases similarly in early training. Motivated by this dynamic, we propose Online Label Refinement (OLR), which progressively corrects potentially noisy labels with majority-voted answers when two conditions hold: a positive slope in the majority answer's rollout pass rate and stable historical consistency across updates, enabling gradual self-correction as the policy improves. We evaluate OLR on six in-distribution mathematical reasoning benchmarks (AIME24/25, AMC, MATH-500, Minerva, and Olympiad) and three out-of-distribution tasks (ARC-c, GPQA-diamond, and MMLU-pro). Across noise ratios from 0.1 to 0.9, OLR consistently improves robustness under both inactive and active noisy-label settings, achieving average gains of 3.6% to 3.9% on in-distribution benchmarks and 3.3% to 4.6% on out-of-distribution evaluations.

cs.LG↗
arXiv2026

Distributional Statistics Restore Training Data Auditability in One-step Distilled Diffusion Models

The proliferation of diffusion models trained on web-scale, provenance-uncertain image collections has made it essential, yet technically unresolved, to determine whether a model has learned from specific copyrighted data without authorization. Current methods primarily rely on the memorization effect, whereby models reconstruct their training images better than unseen ones, to detect unauthorized training data on a per-instance basis. This effect, however, vanishes under distillation, the now-dominant deployment pipeline that compresses compute-intensive teacher diffusion models into efficient {\em student one-step generators} mimicking the teacher's output for real-time user access. As the students train exclusively on teacher-generated outputs and never directly see the teacher's original training data, they carry no per-instance memorization of that upstream data, creating a model laundering loophole that severs the auditable link between a deployed model and its upstream training data. We nonetheless reveal that a distributional memory chain survives under distillation: the student's output distribution remains closer to the teacher's training distribution than to any non-training reference, even if no single training instance is memorized. Exploiting this chain, we develop a distributional unauthorized training data detector, grounded in kernel-based distribution discrepancy, that determines if a candidate dataset of unknown composition is statistically aligned with the student-generated distribution more than held-out non-training datasets, thus tracing provenance back to the teacher's training data. Evaluation across benchmarks and distillation setups confirms reliable detection even when unauthorized data forms a minority of the candidate set, establishing distribution-level auditing as a countermeasure to model laundering and a paradigm for accountable generative AI ecosystems.

cs.LG↗
arXiv2026

UniGame: Turning a Unified Multimodal Model Into Its Own Adversary

Unified Multimodal Models (UMMs) have shown impressive performance in both understanding and generation with a single architecture. However, UMMs still exhibit a fundamental inconsistency: understanding favors compact embeddings, whereas generation favors reconstruction-rich representations. This structural trade-off produces misaligned decision boundaries, degraded cross-modal coherence, and heightened vulnerability under distributional and adversarial shifts. In this paper, we present UniGame, a self-adversarial post-training framework that directly targets the inconsistencies. By applying a lightweight perturber at the shared token interface, UniGame enables the generation branch to actively seek and challenge fragile understanding, turning the model itself into its own adversary. Experiments demonstrate that UniGame significantly improves the consistency (+4.6%). Moreover, it also achieves substantial improvements in understanding (+3.6%), generation (+0.02)on GenEval, out-of-distribution and adversarial robustness (+4.8% and +6.2% on NaturalBench and AdVQA). The framework is architecture-agnostic, introduces less than 1% additional parameters, and is complementary to existing post-training methods. These results position adversarial self-play as a general and effective principle for enhancing the coherence, stability, and unified competence of future multimodal foundation models. The official code is available at: https://github.com/AIFrontierLab/TorchUMM

cs.LG↗
arXiv2026

KnowledgeSmith: Uncovering Knowledge Updating in LLMs with Model Editing and Unlearning

Knowledge editing and machine unlearning are two popular approaches for large language models (LLMs) to stay up-to-date. However, the knowledge updating mechanism of LLMs remains largely unexplored due to insufficient, isolated, and small-scale evaluation. For instance, are LLMs similar to humans in modifying certain knowledge? What differs editing and unlearning as training data increases? This paper proposes KnowledgeSmith, a unified framework to systematically understand the updating mechanism of LLMs. We first cast editing and unlearning as instances of one constrained optimization problem. Then, we propose an automatic dataset generator that provides structured interventions across multiple graph levels and data scales, enabling controlled studies of how different modification strategies propagate through model knowledge. Extensive experiments demonstrate nuanced insights over knowledge propagation, plasticity scaling, consistency, and robustness. For instance, our results show that LLMs do not exhibit similar updating as humans for different levels of knowledge, and there exists consistency-capacity trade-off. We hope our findings can offer suggestions to the design of more reliable and scalable strategies. Code: https://github.com/AIFrontierLab/KnowledgeSmith.git

cs.CL↗
arXiv2026

Learning to Think Fast and Slow for Visual Language Models

When faced with complex problems, we tend to engage in slower, more deliberate thinking. In contrast, for simple questions we give quick, intuitive responses. This dual-system thinking approach allows us to allocate cognitive resources efficiently, reserving deeper analytical effort for tasks that truly require it. However, existing reasoning-oriented visual language models (VLMs) are mostly trained to generate uniformly long reasoning, leading to substantial token waste when concise answers would suffice. In this paper, we observe that pre-trained, general-purpose VLMs manifest variations in response length for different question types, e.g., longer reasoning for math questions while shorter on perception problems. Different from existing work that overrides this prior by stimulating long reasoning without considering the problem complexity, we propose to leverage this prior to develop an explicit dual-mode thinking mechanism. Specifically, we anchor each training instance to either a fast or slow thinking prefix consistent with the model's natural response length tendency. Then, GRPO is adapted to learning dual-system thinking, enabling both autonomous and manual thinking mode selection. Extensive experiments across a wide variety of visual reasoning benchmarks demonstrate that our model, named DualMindVLM, significantly outperforms the base model and achieves state-of-the-art reasoning performance while maintaining high token efficiency.

cs.CV↗
arXiv2026

LAD: Learning Advantage Distribution for Reasoning

Current reinforcement learning objectives for large-model reasoning primarily focus on maximizing expected rewards. This paradigm can lead to overfitting to dominant reward signals, while neglecting alternative yet valid reasoning trajectories, thereby limiting diversity and exploration. To address this issue, we introduce Learning Advantage Distributions (LAD), a distribution-matching framework that replaces advantage maximization with learning the advantage-induced distribution. By establishing the equivalence between the optimal policy update and an advantage-based target distribution, we derive a practical LAD objective formulated as minimizing an $f$-divergence between the policy-induced and advantage-induced distributions. This yields a gradient update that increases likelihood for high-advantage responses while suppressing over-confident probability growth, preventing collapse without requiring auxiliary entropy regularization. LAD incurs no extra training cost compared to GRPO and scales naturally to LLM post-training. In a controlled bandit setting, LAD faithfully recovers the multimodal advantage distribution, validating the theoretical formulation. Experiments on math and code reasoning tasks across several LLM backbones show that LAD reliably improves both accuracy and generative diversity.

cs.LG↗