论文检索

Pinlong Zhao, Zike Ding, Zengshu Ye, Zhou Zhaoting

Multimodal Large Language Models (MLLMs) often suffer from modality bias, where the model disproportionately relies on one modality while neglecting critical information from others. Existing debiasing methods via modality masking create biased responses by completely removing an entire modality, forming an extreme and static training environment. However, real-world multimodal bias often emerges under subtle perturbations (e.g., mild occlusion, noisy instructions), where both modalities are present but the model is tempted to rely on spurious shortcuts. We propose B-APO (Bias-Targeted Adversarial Preference Optimization), which casts debiasing as a bias-targeted min-max game: we generate hard negatives by applying small adversarial perturbations in the latent space to maximally induce language-vision-prior reliance, and then perform preference alignment to enlarge the margin between clean and adversarial responses. This encourages the model to anchor on true cross-modal evidence even under the most adversarial conditions. Extensive experiments on bias and hallucination benchmarks demonstrate that B-APO achieves superior debiasing performance while maintaining general capabilities.

Sriram Padmanabhan, Siyuan Song, Kanishka Misra

Language places subtle constraints on how we make inductive inferences. Developmental evidence by Gelman et al. (2002) has shown children (4 years and older) to differentiate among generic statements ("Bears are daxable"), universally quantified NPs ("all bears are daxable") and indefinite plural NPs ("some bears are daxable") in extending novel properties to a specific member (all > generics > some), suggesting that they represent these types of propositions differently. We test if these subtle differences arise in general purpose statistical learners like Vision Language Models, by replicating the original experiment. On tasking them through a series of precondition tests (robust identification of categories in images and sensitivities to all and some), followed by the original experiment, we find behavioral alignment between models and humans. Post-hoc analyses on their representations revealed that these differences are organized based on inductive constraints and not surface-form differences.

Pinlong Zhao, Huijun Tang, Pengfei Jiao, Mengyang Li

Sampling methods for large language models select candidate tokens based on logit statistics, implicitly assuming that high logits indicate desirable outputs. We identify the Logit Conflation Problem, where a token’s logit aggregates prompt-independent factors, including linguistic fluency and parametric associations, with prompt-relevance. However, only prompt-relevance determines instruction-following quality. We propose SEAL-Sampling (Signal Extraction for Active ReLevance) to isolate this component through attention-weighted attribution. Our framework defines prompt-relevance as the causal effect of prompt content on token logits and establishes attention patterns as an efficient proxy. Experiments on LLaMA-3 demonstrate significant improvements over top-nσ, with gains of 1.8% on AlpacaEval 2.0 and 2.2% on IFEval. Furthermore, attribution scores correlate weakly with raw logits, confirming the extraction of an orthogonal signal. The method is training-free and introduces minimal latency, adding less than 12ms overhead per token.

Tengfei Wen, Xuanang Chen, Ben He, Xiaoliang Cong, Le Sun

Large Language Models (LLMs) struggle with code generation for Ultra Low-Resource Programming Languages (ULRPLs) due to the scarcity of training data. Existing synthetic data generation methods fail in this context, suffering from a severe cold-start problem and resulting in samples that lack diversity. To overcome these challenges, we propose CodeRise, a novel two-stage framework that autonomously generates a high-quality, diverse, and progressively complex curriculum for ULRPLs. The framework first tackles the cold-start and distribution issues by leveraging the full formal syntax of the target language as structural guidance and applying a biased sampling strategy over library modules. Building on this foundation, we fine-tune the model to generate increasingly complex code without explicit syntax input, using an adaptive curriculum and multi-turn self-debugging to progressively improve code quality.We evaluate on two ULRPLs, Tengo and Janet, using migrated HumanEval-Tengo and MBPP-Tengo, as well as our new benchmarks, TengoEval and JanetEval. Experiments show that CodeRise significantly outperforms both training-free and training-based baselines in ultra low-resource environments.

Runxin Cai, Jingtan Wang, Bryan Kian Hsiang Low

In Large Language Model post-training, high-quality data effectively enhances model performance with fine-tuning, highlighting the need to identify high-quality and beneficial fine-tuning data. However, one of the most popular data valuation paradigms, influence function and its variants, are computationally expensive due to their reliance on inverse Hessian-Vector Products (iHVP) computations that scale poorly with increasing model size. To examine whether influence values correlate with efficiently computable intrinsic features, we empirically investigate the distribution of top influential data for the model in fine-tuning, and observe that data with high influence tend to be those with high predictive uncertainty. Yet such highly uncertain samples exhibit a dual nature, which can be either beneficial or detrimental noisy data. Unlike traditional methods that treat uncertainty as a standalone criterion, we introduce a directional indicator to rigorously disentangle these opposing effects. Formally, we propose EULoInf (Entropy-based Uncertainty-aware Lookahead Influence), a computationally efficient valuation framework. By approximating influence via uncertainty and gradient based validation loss lookahead, EULoInf avoids iHVP computation, effectively reducing the iHVP-induced quadratic complexity in model parameters to linear time. We rigorously derive our framework from the influence function. Empirically, it matches or even outperforms prior methods across diverse data valuation tasks and LLM architectures, including mislabel detection and data selection, while reducing computational time and memory usage by over 50%.

Bowen Dong, Wenjun Wang, Xueli Liu, Quanlin Qiu

Understanding whether one event increases or decreases the likelihood of another is critical for real-life applications. Unlike other relationships, promoting and inhibiting relationships capture directional, probabilistic, and context-dependent shifts in event likelihood. A central challenge is to estimate this relative influence from observational data: naive conditional probabilities conflate influence with correlation and are easily distorted by shared contextual confounders. We propose EPIR, a unified framework for estimating promoting and inhibiting relationships from observed event data. EPIR formulates influence as a relative directional effect under comparable contextual conditions, and models event context using : (i) observable history captured and (ii) latent multi-hop propagation mechanisms. EPIR combines context-conditioned predictive evidence with schema-based structural evidence to produce a single signed influence score, where the sign determines promotion versus inhibition. Experiments on real-world datasets show that EPIR outperforms state-of-the-art baselines in accuracy.

Zhixiang Lu, Jionglong Su

Multimodal Large Language Models (MLLMs) in healthcare suffer from severe confirmation bias, often hallucinating visual details to support initial, potentially erroneous diagnostic hypotheses. Existing Chain-of-Thought (CoT) approaches lack intrinsic correction mechanisms, rendering them vulnerable to error propagation. To bridge this gap, we propose Dialectic-Med, a multi-agent framework that enforces diagnostic rigor through adversarial dialectics. Unlike static consensus models, Dialectic-Med orchestrates a dynamic interplay between three role-specialized agents: a proponent that formulates diagnostic hypotheses; an opponent equipped with a novel visual falsification module that actively retrieves contradictory visual evidence to challenge the Proponent; and a mediator that resolves conflicts via a weighted consensus graph. By explicitly modeling the cognitive process of falsification, our framework guarantees that diagnostic reasoning is tightly grounded in verified visual regions. Empirical evaluations on MIMIC-CXR-VQA, VQA-RAD, and PathVQA demonstrate that Dialectic-Med not only achieves state-of-the-art performance but also fundamentally enhances the trustworthiness of the reasoning process. Beyond accuracy, our approach significantly enhances explanation faithfulness and decisively mitigates hallucinations, establishing a new standard over single-agent baselines.

Yifan Simon Liu, Qianfeng Wen, Jiazhou Liang, Mark Zhao, Justin Cui, Anton Korikov, Armin Toroghi, Junyoung Kim, Scott Sanner

Natural Language Recommendation (NLRec) generates item suggestions based on the relevance between user-issued NL requests and NL item description passages. Existing NLRec approaches often use Dense Retrieval (DR) to compute item relevance scores from aggregation of inner products between user request embeddings and relevant passage embeddings. However, DR views the request as the sole relevance label, thus leading to a unimodal scoring function centered on the query embedding that is often a weak proxy for query relevance. To better capture the potential multimodal distribution of the relevance scoring function that may arise from complex NLRec data, we propose **GPR-LLM** that uses Gaussian Process Regression (GPR) with LLM relevance judgments for a subset of candidate passages. Experiments on four NLRec datasets and two LLM backbones demonstrate that GPR-LLM with an RBF kernel, capable of modeling multimodal relevance scoring functions, consistently outperforms simpler unimodal kernels (dot product, cosine similarity), as well as baseline methods including DR, cross-encoder, and pointwise LLM-based relevance scoring by up to 65%. Overall, GPR-LLM provides an efficient and effective approach to NLRec within a minimal LLM labeling budget.

Zhengjun Huang, Zhoujin Tian, Qintian Guo, Fangyuan Zhang, Yingli Zhou, Di Jiang, Zeying Xie, Xiaofang Zhou

Large Language Model (LLM) agents exhibit remarkable conversational and reasoning capabilities but remain constrained by limited context windows and the lack of persistent memory. Recent efforts address these limitations via external memory architectures, often employing graph-based representations, yet most adopt flat, entangled structures that intertwine semantics with topology, leading to redundant representations, unstructured retrieval, and degraded efficiency and accuracy. To resolve these issues, we propose LiCoMemory, an end-to-end agentic memory framework for real-time updating and retrieval, which introduces CogniGraph, a lightweight hierarchical graph that utilizes entities and relations as semantic indexing layers, and employs temporal and hierarchy-aware search with integrated reranking for adaptive and coherent knowledge retrieval. Experiments on long-term dialogue benchmarks, LoCoMo and LongMemEval, show that LiCoMemory not only outperforms established baselines in temporal reasoning, multi-session consistency, and retrieval efficiency, but also notably reduces update latency.

Hieu Man, Ro-ee Tal, Abhishek Kumar, Jaejin Cho, Benjamin Hsu

Multi-hop reasoning remains a fundamental challenge for Retrieval-Augmented Generation (RAG) systems. Recent approaches—from adaptive retrieval to agentic pipelines—struggle to maintain coherent intermediate reasoning states as chains grow longer. We introduce State-Aware RAG, a framework that addresses this limitation through an explicit working memory that serves as a dynamic cognitive workspace for reasoning. Our modular architecture features a lightweight, trainable extractor that learns to actively filter, consolidate, and update this working memory via a novel Path-Outcome Dual Reward paradigm, which balances local coherence with global strategy. The retriever and generator remain frozen, enabling plug-and-play flexibility. Experiments on eight QA benchmarks demonstrate state-of-the-art results, on average achieving +8.6% over the best memory-augmented baseline and +9.3% over the best RL-enhanced baseline. Our architecture generalizes seamlessly to stronger generators and retrievers without retraining, establishing dynamic memory management as a critical yet underexplored dimension for advancing RAG systems.

Qiaoyu Tang, Le Yu, Bowen Yu, Hongyu Lin, Keming Lu, Yaojie Lu, Xianpei Han, Le Sun

Post-training has emerged as a crucial paradigm for adapting large-scale pre-trained models to various tasks, whose effects are fully reflected by delta parameters (i.e., the disparity between post-trained and pre-trained parameters).While numerous studies have explored delta parameter properties via operations like pruning, quantization, low-rank approximation, and extrapolation, a fundamental question remains: what properties of delta parameters are essential for maintaining performance?In this work, we investigate delta parameter properties along two dimensions: magnitude and sign. Through experiments on instruct language models, reasoning language models, and vision models, we find that delta parameters exhibit considerable editability: individual values, distribution shape, relative relationships, and even signs can be substantially modified while maintaining post-trained model’s performance.To understand these phenomena, we propose a loss-based local surrogate analysis that examines editing effects through a second-order Taylor expansion. Our analysis introduces the concept of editing intensity, which helps explain the stability boundaries of different editing operations.

Theodore Glavas, Nikhita Vedula, Dushyanta Dhyani, Yilun Zhu, Shervin Malmasi

Some text generation tasks, such as Attribute Value Extraction (AVE), require decoding multiple independent sequences from the same document context. While standard autoregressive decoding is slow due to its sequential nature, the independence between output sequences offers an opportunity for parallelism. We present Hyper-Parallel Decoding, a novel decoding algorithm that accelerates offline decoding by leveraging both shared memory and computation across batches. HPD enables out-of-order token generation through position ID manipulation, significantly improving efficiency. Experiments on AVE show that attribute-value pairs are conditionally independent, enabling us to parallelize value generation within each prompt. By further stacking multiple documents within a single prompt, we can decode in parallel up to 96 tokens per prompt. HPD works with all LLMs, and reduces both inference costs and total inference time by up to 13.8X without compromising output quality, potentially saving hundreds of thousands of dollars on industry AVE tasks. Although designed for attribute extraction, HPD makes no assumptions unique to the AVE domain and can in theory be applied to other scenarios with independent output structures.

Yi Feng, Zijie Yang, Chen Zhang, Wenxuan Zhang, Dongming Zhang, Liping Jing, Jian Yu

Existing psychological counseling datasets often suffer from monolithic client personas, insufficient therapeutic depth, and a lack of process controllability. To address these critical limitations, we propose PsyChain, a chain-of-agents framework that evolves static counseling corpora into high-fidelity dialogues through collaborative simulation which explicitly models client personality, stage progression, safety monitoring, and expert supervision. PsyChain involves a Client Profiler that extracts life scenarios and pairs them with psychological personality archetypes to synthesize diverse profiles.To simulate the complete counseling process, five specialized agents—Process Monitor, Client Speaker, Safety Monitor, Counselor Supervisor, and Counselor Speaker—collaborate and interact autonomously at each dialogue turn to ensure therapeutic professionalism and safety.We apply this to construct PsyChainD, a Chinese dataset of 10,456 dialogues featuring systematically diverse client profiles. Extensive evaluation across client side, counselor side and overall quality shows substantial improvements. The model trained on PsyChainD achieves 61-91% win rates against domain-specific baselines in pairwise evaluation and the highest average score in human evaluation, indicating potential for real-world counseling.

Zhi Rui Tam, Yung-Yu Shih, Yen-Wei Lee, Ya-Ting Pai, Wen Yu Chang, Yun-Nung Chen

Vision-Language Models (VLMs) often struggle in Taiwanese Mandarin environments due to region-specific orthographic and cultural context. We introduce VisTW, a comprehensive benchmark featuring (i) multiple-choice questions (3,795 academic questions) and (ii) free-form generation evaluation (141 Taiwanese-context free-form pairs). Beyond standard accuracy, we investigate character mixing— the unintended production of Simplified Chinese characters under Taiwanese-Mandarin-style prompts—and propose a human-grounded purity penalty derived from perceptual thresholds measured from users. Our evaluation reveals substantial character contamination (3%–19%) across state-of-the-art VLMs. We find that Gemini-3-Pro significantly outperforms the strongest open-weight baseline, Qwen3 235B MoE, by up to 22 percentage points on dialogue tasks once the purity penalty is applied. These results highlight orthographic consistency as a vital, yet overlooked, dimension for localized multimodal evaluation and deployment.

Dhruv Madhwal, Lyuxin David Zhang, Dan Roth, Tomer Wolfson, Vivek Gupta

Large language models often struggle to recognize their knowledge limits in closed-book question answering, leading to confident hallucinations. While decomposed prompting is typically used to improve accuracy, we investigate its impact on reliability. We evaluate three task-equivalent prompting regimes: Direct, Assistive, and Incremental, across different model scales and multi-hop QA benchmarks. We find that although accuracy gains from decomposition diminish in frontier models, disagreements between prompting regimes remain highly indicative of potential errors. Because factual knowledge is typically stable while hallucinations are stochastic, cross-regime agreement provides a precise signal of internal uncertainty. We leverage this signal to implement a training-free abstention policy that requires no retrieval or fine-tuning. Our results show that disagreement-based abstention outperforms standard uncertainty baselines as an error detector, improving both F1 and AUROC across settings. This demonstrates that decomposition-based prompting can serve as a practical diagnostic probe for model reliability in closed-book QA.

Seyedarmin Azizi, Erfan Baghaei Potraghloo, Souvik Kundu, Massoud Pedram

Large language models (LLMs) demonstrate strong performance on multi-step reasoning tasks by producing intermediate explanations, commonly referred to as chains of thought (CoTs). However, the generated rationales are typically verbose, consuming many additional tokens, and thus degrading throughput and increasing inference energy consumption. Interestingly, we find that verbose and concise CoTs correspond to distinct regions in the model’s intermediate activation space, suggesting that verbosity is a steerable latent attribute. Building on this observation, we develop an inference-time method to automatically steer the model response towards concise reasoning traces without updating model parameters. Our method, dubbed _ASC_ (Activation-Steered Compression), generates concise CoTs by directly adjusting internal representations via activation steering. A key component of ASC is **Contrastive Energy-Based Steering (CES)**, a principled procedure to learn a _single_ steering vector from a small set of verbose–concise CoT pairs by optimizing a length-normalized contrastive energy objective. To further ensure reliable steering and preserve general utility, CES enforces a differentiable **KL trust region** during steering vector optimization, explicitly constraining the distribution shift within a specified budget. With only 100 pairs of verbose–concise examples, ASC reduces the generated token length by as much as 69.4% across five reasoning benchmarks (MATH500, GSM8K, LiveCodeBench, GSM8K-Hard, and AQuA-RAT) while maintaining accuracy across models with 1.5B, 7B, 8B, and 32B parameters. On MATH500, ASC achieves an end-to-end inference speed-up of 2.7× on an 8B model.

Liuliu Chen, Elise Carrotte, Brian E. Chapman, Jo Robinson, Mike Conway

Suicide memes are memes used to express suicide-related thoughts or comment on suicide-related issues. Suicide memes are increasingly common on social media, yet remain poorly understood and potentially harmful. There is an urgent need to better understand their characteristics and to develop appropriate content moderation strategies that limits users’ exposure to potentially harmful content. Currently, the absence of annotated datasets of suicide memes remains a key barrier to developing and evaluating automated moderation approaches. In this paper, we introduce FigSIM, the first dataset designed for fine-grained analysis of suicide memes. The dataset consists of 1049 memes, each annotated for (1) fine-grained suicide severity levels, (2) figurative phenomena (e.g. metaphors), and (3) suicide-related content (e.g. suicide method depiction). We benchmark 16 unimodal and multimodal models across three tasks: figurative language, suicide severity, and suicide-related content detection. Overall, FigSIM demonstrates that suicide memes pose unique challenges for both modeling and content moderation. Analysis revealed biases, such as underprediction of higher suicide severity levels, especially for figurative memes.

Yue Guo, Fanfu Wang, Jianwei Lv, Xincheng Shi, Yuchen Li, Youya Wang, Yunsheng Zeng, Yujing Liu, Yunhao Qiao, Gen Li 等

Clinical Decision Support Systems (CDSSs) provide reasoning and inquiry guidance for physicians, yet they face notable challenges, including high maintenance costs and low generalization capability.Recently, Large Language Models (LLMs) have been widely adopted in healthcare due to their extensive knowledge reserves, retrieval, and communication capabilities. While LLMs show promise and excel at medical benchmarks, their diagnostic reasoning and inquiry skills are constrained.To mitigate this issue, we propose (1) Clinical Diagnostic Reasoning Data (CDRD) structure to capture abstract clinical reasoning logic, and a pipeline for its construction, and (2) the Dr. Assistant, a clinical diagnostic model equipped with clinical reasoning and inquiry skills. Its training involves a two-stage process: SFT, followed by RL with a tailored reward function.We also introduce a benchmark to evaluate both diagnostic reasoning and inquiry.Our experiments demonstrate that the Dr. Assistant outperforms open-source models and achieves competitive performance to closed-source models, providing an effective solution for clinical diagnostic inquiry guidance. Project information can be found at: https://github.com/YGswu/Dr.-Assistant.

Nitish Shukla, Surgan Jandial, Arun Ross

Vision-Language Models (VLMs) have demonstrated remarkable progress in single-image understanding, yet effective reasoning across multiple images remains challenging. We identify a critical capability gap in existing multi-image alignment approaches: current methods focus primarily on localized reasoning with pre-specified image indices (“Look at Image 3 and...”), bypassing the essential skills of global visual search and autonomous cross-image comparison. To address this limitation, we introduce a Simple-to-Hard (S2H) learning framework that systematically constructs multi-image preference data across three hierarchical reasoning levels requiring an increasing level of capabilities: (1) single-image localized reasoning, (2) multi-image localized comparison, and (3) global visual search. Unlike prior work that relies on model-specific attributes, such as hallucinations or attention heuristics, to generate preference pairs, our approach leverages prompt-driven complexity to create chosen/rejected pairs that are applicable across different models. Through extensive evaluations on LLaVA and Qwen-VL models, we show that our diverse multi-image reasoning data significantly enhances multi-image reasoning performance, yielding significant improvements over baseline methods across benchmarks. Importantly, our approach maintains strong single-image reasoning performance while simultaneously strengthening multi-image understanding capabilities, thus advancing the state of the art for holistic visual preference alignment.

Mingkuan Zhao, Yide Gao, Wentao Hu, Suquan Chen, Tianchen Huang, Zhenhua An, Zetao Chang, Xiayu Sun, Yuheng Min

Large Language Models (LLMs) frequently exhibit “contextual disregard” when faced with input evidence that conflicts with their internal parametric memory, leading to persistent factual hallucinations. Existing mitigation strategies primarily rely on suppressing specific neuron activations or employing computationally expensive contrastive decoding mechanisms, which often result in increased perplexity or significantly elevated inference latency. To address these limitations, we propose Resonant Context Anchoring (RCA), a lightweight inference-time intervention method grounded in the perspective of residual stream signal dynamics. RCA aims to resolve the signal attenuation of external evidence during its propagation through deep networks. The core mechanism involves the orthogonal decoupling of routing logic and information magnitude within the self-attention module. By utilizing raw pre-softmax attention scores as an instantaneous metric of semantic alignment, we construct a dynamic gain field via non-linear rectification to selectively amplify the norms of value vectors corresponding to context tokens, without altering the attention probability distribution. This mechanism effectively elevates the signal-to-noise ratio (SNR) of input evidence within the residual stream mixture, thereby robustly anchoring the generation trajectory to the truthful context during inference. Extensive experiments on the Llama-3 model series demonstrate that RCA significantly improves contextual faithfulness across multiple factual consistency and strong knowledge-conflict tasks, effectively suppressing parametric hallucinations. Furthermore, results confirm that as a training-free and computationally negligible plug-and-play module, RCA achieves a Pareto improvement in faithfulness and fluency while maintaining the model’s general language understanding capabilities. Our code is available at https://anonymous.4open.science/r/RCA-Implementation-D8B5