Advances in Multimodal Large Language Models (MLLMs) intensify concerns about data safety, making Machine Unlearning (MU), the selective removal of harmful/private information, a critical necessity. However, existing MU benchmarks for MLLMs are limited by a lack of image diversity, coarse-grained unlearning target, and insufficient evaluation scenarios, which fail to capture the complexity of real-world applications. To facilitate the development of MLLMs unlearning and alleviate the aforementioned limitations, we introduce OFFSIDE, a novel benchmark for evaluating misinformation unlearning in MLLMs. This manually curated dataset contains 15.68K records for 80 players, providing a comprehensive framework with four test sets to assess forgetting efficacy, generalization, utility, and robustness. OFFSIDE supports advanced unlearning targets, such as fine-grained unlearning and visual rumor removal. Our extensive evaluation of multiple baselines not only extends key findings from LLM MU to MLLM MU: (1) unlearned rumors can be easily recovered through relearning and (2) all methods are vulnerable to prompt attacks, but also introduces novel insights in the context of MLLM: (1) unimodal methods fail to handle multimodal rumors, (2) unlearning efficacy is primarily driven by catastrophic forgetting statistically, and (3) all methods struggle with visual rumors (rumors embedded in images). These results expose significant vulnerabilities in current approaches, highlighting the need for more robust multimodal unlearning solutions.
论文检索
输入标题、作者或关键词，从 2,245 篇学术成果中精准定位
Reinforcement Learning with Verifiable Rewards (RLVR) is an essential paradigm that enhances the reasoning capabilities of Large Language Models (LLMs). However, existing methods typically rely on static policy optimization schemes that misalign with the model’s evolving reasoning capabilities. To address this issue, we propose Adaptive Power-Mean Policy Optimization (APMPO), which comprises two main innovations: Power-Mean Policy Optimization (PMPO) and Feedback-Adaptive Clipping (FAC). Specifically, PMPO introduces a generalized power-mean objective. This enables the model to adaptively transition from the signal-amplifying behavior of the arithmetic mean to the consistency-enforcing behavior of the geometric mean. FAC adaptively adjusts clipping bounds based on real-time reward statistics to overcome the limitations of static mechanisms. Capitalizing on these innovations, APMPO improves learning dynamics and reasoning performance. Extensive experiments on nine datasets across three reasoning tasks showcase the superiority of APMPO over state-of-the-art RLVR-based baselines. For instance, APMPO boosts the average Pass@1 score on mathematical reasoning benchmarks by 3.0 points compared to GRPO when using Qwen2.5-3B-Instruct.
Preference optimization is fundamental for aligning large language models. While existing methods use sample weighting, they typically rely on static functions of instantaneous model states and ignore temporal learning dynamics. We contend that a sample’s value evolves throughout training, characterized by patterns such as stable convergence or noisy oscillation. We propose MetaPO, a framework that meta-learns adaptive weights using three temporal features: reward margin evolution, learning volatility, and reference deviation. Through bilevel optimization on validation data, MetaPO automatically discovers weighting strategies tailored to specific datasets. Experiments on models ranging from 7B to 70B parameters demonstrate statistically significant improvements over strong baselines, achieving gains of up to 2.4 points on AlpacaEval 2.0 and Arena-Hard. Interpretability analysis confirms that temporal features drive over 70% of the weighting decisions and that the learned weights correlate strongly with sample quality.
Large Language Models (LLMs) face information overload when handling long contexts, particularly in Retrieval-Augmented Generation (RAG) where extensive supporting documents introduce redundant content that interferes with reasoning. Context engineering has emerged to address these challenges, yet existing methods rely on lexical or token-level features that fragment semantic units and fail to capture conceptually essential content. We propose an unsupervised context compression framework leveraging Abstract Meaning Representation (AMR) to preserve semantically essential information while filtering irrelevant text. By quantifying node-level entropy within AMR graphs, our method estimates the conceptual importance of each node, enabling retention of core semantics. Specifically, we construct AMR graphs from retrieved contexts, compute the conceptual entropy of each node, and identify statistically significant concepts to form a condensed, semantically focused context. Experiments on the PopQA and EntityQuestions datasets demonstrate that our method outperforms vanilla RAG and existing baselines, achieving superior accuracy while substantially reducing context length. To the best of our knowledge, this is the first work introducing AMR-based conceptual entropy for context compression, demonstrating the potential of structured linguistic representations in context engineering.
Large language models (LLMs) have demonstrated significant utility in real-world applications, exhibiting impressive capabilities in natural language processing and understanding. Benchmark evaluations are crucial for assessing the capabilities of LLMs as they can provide a comprehensive assessment of their strengths and weaknesses. However, current evaluation methods often overlook the inherent randomness of LLMs by employing deterministic generation strategies or relying on a single random sample, resulting in unaccounted sampling variance and unreliable benchmark score estimates. In this paper, we propose a hierarchical statistical model that provides a more comprehensive representation of the benchmarking process by incorporating both benchmark characteristics and LLM randomness. We show that leveraging multiple generations improves the accuracy of estimating the benchmark score and reduces variance. Multiple generations also allow us to define \mathbb{P} (correct), a prompt-level difficulty score based on correct ratios, providing fine-grained insights into individual prompts. Additionally, we create a data map that visualizes difficulty and semantics of prompts, enabling error detection and quality control in benchmark construction.
The slow thinking paradigm has been widely validated to enhance the reasoning capabilities of Large Language Models (LLMs), but it introduces notable reasoning inefficiencies: models often overthink simple tasks while prematurely shifting their reasoning paths when addressing complex problems. To address this, we propose AdapThink, a simple yet efficient framework for adaptive reasoning preference control. Unlike methods imposing uniform length constraints, AdapThink dynamically adjusts reflection preferences based on group-level distributional statistics of reasoning length and reflection intensity. We further introduce a dispersion-based diversity sampling mechanism that maximizes the geometric spread of reasoning patterns, accelerating learning through exposure to diverse problem-solving strategies. Across mathematical reasoning and code generation benchmarks, AdapThink reduces average response length by 17.1%-21.4% while improving performance by 6.12-6.59 points under 32K token budgets, demonstrating superior efficiency and robustness against reward hacking compared to strong baselines.
Robust Uncertainty Quantification for Self-Evolving Large Language Models via Continual Domain Pretraining
PDF ↗Continual Learning (CL) is essential for enabling self-evolving large language models (LLMs) to adapt and remain effective amid rapid knowledge growth. Yet, despite its importance, little attention has been given to establishing statistical reliability guarantees for LLMs under CL, particularly in the setting of continual domain pretraining (CDP). Conformal Prediction (CP) has shown promise in offering correctness guarantees for LLMs, but it faces major challenges in CDP: testing data often stems from unknown or shifting domain distributions, under which CP may no longer provide valid guarantees. Moreover, when high coverage is required, CP can yield excessively large prediction sets for unanswerable queries, reducing informativeness. To address these challenges, we introduce an adaptive rejection and non-exchangeable CP framework. Our method first estimates the distribution of questions across domains in the test set using transformer-based clustering, then reweights or resamples the calibration data accordingly. Building on this, adaptive rejection CP allows the LLM to selectively abstain from answering when its confidence or competence shifts significantly. Extensive experiments demonstrate that our framework enhances both the effectiveness and reliability of CP under CDP scenarios. Our code is available at: https://github.com/AlearZhou/CPCL
MCLE-Mol: Empowering LLM with Molecular Comprehension and Low-Cost Continual Evolution for Interpretable Property Prediction
PDF ↗Large language models (LLMs) offer a new paradigm for molecular property prediction (MPP), yet a semantic gap between natural language and molecular representations limits LLMs’ ability to capture structure–activity relationships (SAR). Recent approaches have explored injecting structure-level information into LLMs, primarily modeling associations based on statistical regularities. However, these methods are prone to misinterpreting coincidental associations as general principles, imposing a bottleneck on predictive performance. To tackle the challenges above, we propose MCLE-Mol, an ML–LLM–Rule collaborative framework for MPP. It bridges the semantic gap by injecting ML-derived substructure attribution values into LLMs, utilizing Context-Calibrated Substructure Attribution Rules (CCSAR) to calibrate these attributions under specific chemical contexts to mitigate such misinterpretation. In addition, MCLE-Mol introduces a low-cost continual evolution strategy that updates CCSAR with frozen model parameters to adapt to dynamic chemical spaces. Experiments on multiple benchmark datasets demonstrate that MCLE-Mol outperforms all baselines, successfully resolving the trade-off between predictive performance and interpretability.
Event prediction plays a critical role in high-stakes applications such as military operations, public safety, and healthcare. Current methods learn temporal knowledge graphs to predict events at future timestamps, and the predictions directly influence decision-making and resource allocation. However, these methods lack rigorous uncertainty quantification, which limits their reliability for decision-making, especially in high-stakes scenarios where the cost of errors is high. In this paper, we propose CFEP, a conformal prediction framework tailored for event prediction to address this challenge. This is achieved through end-to-end optimization that ensures coverage while improving efficiency. Specifically, we first introduce non-conformity score diffusion, which captures both topological and temporal uncertainty in temporal knowledge graphs. Additionally, we propose an efficiency-aware optimization algorithm to reduce the coverage gap and improve computational efficiency. Experimental results on three public datasets demonstrate that our approach consistently guarantees statistical coverage while improving efficiency. The code and datasets are available at https://github.com/hucheng-IIE/CFEP.
Verifiable Parameterization of Bayesian Networks from Scientific Literature: Unlocking Unstructured Empirical Evidence
PDF ↗Learning Bayesian Networks typically requires access to raw tabular data to estimate conditional probabilities. However, in many scientific domains, raw data is unavailable due to privacy concerns or general lack of access, while structured statistical summaries are increasingly accessible through large language models and published literature. We propose and evaluate five distinct strategies to reconstruct local conditional probability tables solely from statistical summaries in order to parameterize Bayesian Networks. Our comprehensive evaluation across mixed-type synthetic networks demonstrates that copula-based methods significantly outperform standard baselines, offering a viable path for knowledge integration from heterogeneous sources – unlocking the wealth of published knowledge for causal modeling while ensuring transparency and verifiability.
WavAlign: Enhancing Intelligence and Expressiveness in Spoken Dialogue Models via Adaptive Hybrid Post-Training
PDF ↗End-to-end spoken dialogue models have garnered significant attention because they offer a higher potential ceiling in expressiveness and perceptual ability than cascaded systems. However, the intelligence and expressiveness of current open-source spoken dialogue models often remain below expectations. Motivated by the success of online reinforcement learning(RL) in other domains, one might attempt to directly apply preference optimization to spoken dialogue models, yet this transfer is non-trivial. We analyze these obstacles from the perspectives of reward modeling and rollout sampling, focusing on how sparse preference supervision interacts with dense speech generation under shared-parameter updates. Based on the analysis, we propose a modality-aware adaptive post-training recipe that makes RL practical for spoken dialogue: it constrains preference updates to the semantic channel and improves acoustic behavior via explicit anchoring, while dynamically regulating their mixture from rollout statistics to avoid unreliable preference gradients. We evaluate the method across multiple spoken dialogue benchmarks and representative architectures, and observe consistent improvements in semantic quality and speech expressiveness.
BV-Blend: Uncertainty-Weighted Historical Baselines for Stable Critic-Free RL with Verifiable Rewards
PDF ↗Critic-free reinforcement learning with verifiable rewards (RLVR), exemplified by Group Relative Policy Optimization (GRPO), avoids training a value function (critic) and reduces memory and compute overhead relative to critic-based PPO pipelines for aligning large language models. However, GRPO-style advantage estimation depends on prompt-local (within-prompt-group) reward statistics and can be unstable. In particular, when all rollouts in a prompt group receive identical rewards, the within-group reward variance becomes zero, and group normalization yields *zero* advantages for that group, impeding learning in cold-start regimes with binary verifiers. We introduce **BV-Blend**, a critic-free framework that stabilizes advantage estimation by combining prompt-local on-policy statistics with semantic-cluster-conditioned historical moments. BV-Blend maintains EMA-tracked reward moments for each cluster, derives a confidence weight from a standard error of the mean (SEM) proxy, and uses this weight to blend historical and prompt-local baseline and variance statistics into a standardized advantage for PPO-style clipped updates. Experiments on verifiable reasoning benchmarks show that BV-Blend improves training stability and performance, and remains robust in regimes where group-normalized methods may stall.
Do LLMs Really Know What They Don’t Know? Internal States Mainly Reflect Knowledge Recall Rather Than Truthfulness
PDF ↗Recent work suggests that LLMs "know what they don’t know", positing that hallucinated and factually correct outputs arise from distinct internal processes and can therefore be distinguished using internal signals.However, hallucinations have multifaceted causes: beyond simple knowledge gaps, they can emerge from training incentives that encourage models to exploit statistical shortcuts or spurious associations learned during pretraining.In this paper, we argue that when LLMs rely on such learned associations to produce hallucinations, their internal processes are mechanistically similar to those of factual recall, as both stem from strong statistical correlations encoded in the model’s parameters.To verify this, we propose a novel taxonomy categorizing hallucinations into Unassociated Hallucinations (UHs), where outputs lack parametric grounding, and Associated Hallucinations (AHs), which are driven by spurious associations. Through mechanistic analysis, we compare their computational processes and hidden-state geometries with factually correct outputs.Our results show that hidden states primarily reflect whether the model is recalling parametric knowledge rather than the truthfulness of the output itself. Consequently, AHs exhibit hidden-state geometries that largely overlap with factual outputs, rendering standard detection methods ineffective. In contrast, UHs exhibit distinctive, clustered representations that facilitate reliable detection.
Video-LLMs face a fundamental tension in long-video reasoning: static, sparse frame sampling either dilutes evidence across task-irrelevant segments at significant cost or misses fine-grained temporal semantics altogether. We propose a novel, cognitively-inspired task — Endomorphic Multimodal Compression (EMC) — as a structurally-constrained sufficient-statistic problem for VideoQA, and formulate it as an endomorphic transformation F_EMC : (V, Q) → (v, q) that compresses the multimodal input while preserving answer invariance across reasonable downstream models. The endomorphic form keeps the compressed output in the downstream pipeline’s native task space — a structural mirror of the filter-then-reason mechanism in the cognitive literature motivating EMC — distinguishing it from latent-code compression (IB / VIB) and making the formulation extensible to other multimodal settings. Under the Markov chain A → (V, Q) → (v, q), EMC realizes the classical sufficiency condition I((v, q); A) = I((V, Q); A) in its VideoQA-natural form. As a modular front-end, EMC plugs into both Video Instruction Tuning and Video Question Answering pipelines. We release the first dedicated benchmark and propose ReSimplifyIt, an EMC baseline surpassing prior methods by 0.40 F-1 with competitive query rewriting. Integrating EMC yields relative gains of 7.33% in training and 33.7% in inference for video-language understanding.
This paper introduces Late Code Chunking (LC^2), a chunking strategy designed to improve the semantic understanding of code segments for Large Language Models (LLMs). Repository-level code completion requires predicting the completion of unfinished code by leveraging cross-file context spread across a repository. However, when retrieved fragments have missing semantics—the loss of structural or behavioral information during chunking—LLMs struggle to interpret the target code. To address this, LC^2 refines retrieved chunks by constructing a dual context: a "Code Retrieval Context" optimized for similarity-based search, and a "Code Comprehension Context" that serves as a late enrichment step through context expansion and augmentation. This dual-context design reduces information loss due to chunking and enhances the ability of LLMs to utilize retrieved code. Additionally, we introduce an Asymmetric Query-Chunk Sizing strategy to further optimize retrieval quality by minimizing query noise. Our experiments demonstrate that LC^2 provides robust performance gains, achieving a statistically significant 19.7% improvement in Exact Match accuracy on the CrossCodeEval benchmark compared to the best existing chunking method.
Membership inference attacks (MIAs) are a canonical way to assess a machine learning model’s privacy properties. Although several attempts have been made to evaluate MIAs on language models, the extant literature has suffered numerous difficulties in constructing clean evaluations to test new techniques. In particular, subtle distribution shifts between member and non-member sets can undermine the statistical validity of MIAs; recent work has underscored this by showing that “blind” methods with no access to the underlying model can perform far better than published methods on the same benchmarks. This paper constructs a benchmark for principled evaluation of MIAs against LLMs, by leveraging the insight that training data before and after a fixed point during training are drawn from the same distribution. Therefore, all open-source models with intermediate checkpoints and public training data can be converted into MIA testbeds. We apply our framework to a half-dozen published attacks on the Pythia and OLMo family of models, from 70M to 7B parameters. To facilitate further privacy research, we open-source a modular library for designing and implementing attacks in this setting: https://github.com/safr-ai-lab/pandora_llm.
Measuring the **breadth** of a word’s meaning, or its spread across contexts, has become feasible with contextualized token embeddings. A word type can be represented as a cloud of token vectors, with dispersion-based statistics serving as proxies for contextual diversity (Nagata and Tanaka-Ishii, ACL2025). These measurements are useful for deciding appropriate sense distinctions when constructing thesauri and domain-specific dictionaries. However, when comparing the breadth of two word types, naive hypothesis testing on dispersion can be misleading: differences in semantic direction can masquerade as dispersion differences, inflating Type-I error and yielding “statistically significant” outcomes even when there is no true breadth difference. This is problematic because significance testing should distinguish genuine effects from incidental fluctuations in small-difference regimes. We propose a Householder-aligned permutation test to isolate dispersion differences from directional differences. Our method applies a single Householder reflection to align the mean directions of the two word types and then performs a permutation test on the aligned token clouds, yielding calibrated, non-parametric p-values. For practicality, we introduce a GPU-oriented implementation that batches permutations and linear algebra operations. **Empirically, our alignment reduced Type-I error by 32.5% while preserving sensitivity to genuine breadth differences, and achieved a 23\times speedup over the CPU baseline.**
Causal reasoning in natural language requires identifying relevant variables, understanding their interactions, and reasoning about effects and interventions, often under noisy or ambiguous conditions. While large language models (LLMs) exhibit strong general reasoning abilities, they struggle to disentangle correlation from causation, particularly when observations are partially incorrect or irrelevant information is present. In this work, we introduce NoisyCausal, a new benchmark designed to evaluate causal reasoning under structured noise. Each instance is generated from a ground-truth causal graph and contextualized with a natural language scenario by injecting controllable forms of noise, such as irrelevant distractors, value perturbations, confounding, and partial observability. Moreover, we propose a modular reasoning framework that combines LLMs with explicit causal structure to address these challenges. Our method prompts the LLM to extract variables, construct a causal graph from context, and then reformulates the reasoning task as a structured prompt grounded in this graph. Rather than relying on statistical patterns alone, the LLM is guided by symbolic structure, enabling more interpretable and robust inference. Experimental results show that our method significantly outperforms standard prompting and reasoning baselines on NoisyCausal. Furthermore, it generalizes well to external benchmarks such as Cladder without task-specific tuning. Our findings highlight the importance of combining causal abstractions with language-driven reasoning to achieve faithful and robust causal understanding in LLMs.
Detecting LLM-generated code is crucial for ensuring software provenance, security, reliability, and licensing compliance. Existing training-free detectors, mostly adapted from text-based methods, rely on global statistics of the Token Perplexity Sequence (TPS) and struggle with code. We reveal a key insight: despite the convergence of global statistics, LLM-generated and human-written code differ fundamentally in their local TPS dynamics: the former shows narrow transient spikes while the latter exhibits broad sustained fluctuations. To capture this distinction, we introduce CodeRipple, a novel training-free detection framework that employs wavelet analysis to characterize TPS morphology across scales. It jointly leverages the Stationary Wavelet Transform to model fluctuation shape and the Discrete Wavelet Transform to quantify cross-scale energy distribution. Evaluated on three challenging benchmarks spanning diverse programming languages, multiple generating LLMs, and various evasion strategies, CodeRipple consistently outperforms existing training-free methods, demonstrating its superior effectiveness and generalizability without any model training. Code available at: https://github.com/yaoxingyu77/CodeRipple.
The widespread deployment of Large Language Models (LLMs) has spurred significant progress in the detection of LLM-generated text. However, existing detection methods often rely on statistical features that are insufficient for reliable detection; for example, even though LLM-generated and human-written texts exhibit different probability distributions in surrogate models, they can produce nearly identical entropy values, thereby conflating the two types of text. In this paper, we propose that modulating the decoding temperature and monitoring how the probability distributions respond can better probe the intrinsic discrepancies between two types of text. Building upon this insight, we introduce a new feature termed Temperature Sensitivity (TS) and demonstrate that LLM-generated text tends to exhibit higher TS than human-written text. Finally, we propose NTS, a novel and simple zero-shot detector built upon normalized temperature sensitivity. Extensive experiments across three datasets, multiple domains, and various source models demonstrate the superior effectiveness and robustness of our proposed approach. Code avaliable at: https://github.com/Shixuan-Ma/NTS.