Large Language Models (LLMs) face a fundamental challenge with delayed disambiguation: **How is the meaning of an ambiguous word updated when clarifying context arrives only after it has been processed?** While LLMs possess the latent capacity to resolve such ambiguities—as revealed when a full, non-causal context is provided—their unidirectional architecture prevents immediate updates. We investigate the underlying computational mechanism and show this semantic re-evaluation is deferred to subsequent tokens in a process we term "Deferred Semantic Drift (DSD)". Through targeted analysis of attentional pathways, we find that later tokens actively retrieve context-dependent "informational packets" from the ambiguous word’s value vector to steer the final interpretation. We demonstrate this mechanism in metaphor comprehension and provide causal validation by steering model outputs towards literal or metaphorical meanings via targeted activation interventions. This research uncovers a key computational strategy for meaning construction, offering crucial insights for understanding and guiding the behavior of LLMs.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
With the widespread proliferation of the Internet, the spread of fake news has accelerated significantly, evolving from single-text content to multimodal forms that include images and videos. The task of Multimodal Fake News Detection (MFND) takes both text and relevant images as input for fake news identification. However, issues such as image noise and inaccurate focus of visual features often lead to insufficient attention to critical information within images during multimodal fusion. To effectively address these challenges, we propose a covariance matrix-driven image channel allocation method. This method first expands the number of original channel maps, then evaluates the importance of image channels through the covariance matrix and assigns importance scores to the expanded channel maps, thereby redirecting the focus of visual features. Subsequently, we design a multimodal fusion strategy based on a multilayer co-attention mechanism to achieve dynamic fusion across modalities. Finally, a contrastive learning loss is introduced to enhance the alignment between textual and visual modalities. Extensive experiments demonstrate that our method achieves state-of-the-art performance on three public multimodal fake news detection benchmark datasets.
Multi-Scale Spectral Selection and Entropy-Guided Uncertainty Fusion for Multimodal Rumor Detection
PDF ↗Multimodal content combining textual and visual information poses significant challenges for rumor detection on social media. Compared to traditional spatial domain features, frequency domain features have attracted increasing attention due to their stronger discriminative capabilities. However, existing methods still fall short in capturing cross-modal semantic inconsistencies and often overlook inherent noise in multimodal features, which limits overall detection performance. To address these issues, we propose a novel multimodal rumor detection method based on multi-scale spectral selection and entropy-guided uncertainty fusion. Specifically, we first apply the Discrete Cosine Transform (DCT) to image and text features to convert them into the frequency domain. Then, multi-scale convolutional filters are employed to extract fine-grained information across different frequency scales. Next, modality separation is performed to capture both shared and modality-specific features, enabling more effective cross-modal representation learning. Finally, entropy is used to estimate the uncertainty of each prediction branch, calculate confidence scores, and perform adaptive weighted fusion accordingly. Experimental results on multiple benchmark datasets demonstrate that our method outperforms existing state-of-the-art approaches in multimodal rumor detection, demonstrating stronger detection capability and robustness.
Hybrid-Vector Retrieval for Visually Rich Documents: Combining Single-Vector Efficiency and Multi-Vector Accuracy
PDF ↗Retrieval over visually rich documents is essential for tasks such as legal discovery, scientific search, and enterprise knowledge management. Existing approaches fall into two paradigms: single-vector retrieval, which is efficient but coarse, and multi-vector retrieval, which is accurate but computationally expensive. To address this trade-off, we propose HEAVEN, a plug-and-play two-stage hybrid-vector framework. In the first stage, HEAVEN efficiently retrieves candidate pages using a single-vector method over Visually-Summarized Pages (VS-Pages), which assemble representative visual layouts from multiple pages. In the second stage, it reranks candidates with a multi-vector method while filtering query tokens by linguistic importance to reduce redundant computations. To evaluate retrieval systems under realistic conditions, we also introduce ViMDoc, a benchmark for visually rich, multi-document, and long-document retrieval. Across four benchmarks, HEAVEN attains 99.87% of the Recall@1 performance of multi-vector models on average while reducing per-query computation by 99.82%, achieving efficiency and accuracy. Our code and datasets are available at: https://github.com/juyeonnn/HEAVEN
The rise of reasoning models necessitates large-scale verifiable data, for which programming tasks serve as an ideal source. However, while competitive programming platforms provide abundant problems and solutions, high-quality test cases for verification remain scarce. Existing approaches attempt to synthesize test cases using Large Language Models (LLMs), but rely solely on the model’s intrinsic generation capabilities without external feedback, frequently resulting in insufficiently diverse cases. To address this limitation, we propose a \textbf{Feedback-Driven Iterative Framework} for comprehensive test case construction. Specifically, our method leverages the LLM to generate initial test cases, executes them against known correct and incorrect solutions, and utilizes the failed results as feedback to guide the LLM in refining the test cases toward high fidelity and discriminability. We then apply this method to the CodeContests dataset to construct an optimized high-quality derivative, \textbf{CodeContests-O}. Evaluating against the entire pool of solutions (1.1 \times 10^7 in total), our dataset achieves an average True Positive Rate (TPR) of 89.35% and True Negative Rate (TNR) of 90.30%, significantly outperforming the CodeContests and CodeContests+ by margins of 4.30% and 8.78%, respectively. Furthermore, fine-tuning the Qwen2.5-7B model on CodeContests-O results in a 9.52% improvement on LiveCodeBench (Pass@1). Experiments demonstrate the effectiveness of our framework and the quality of CodeContests-O.
Although LLM-based agents are proven to master tool orchestration in scientific fields, particularly chemistry, their single-task performance remains limited by underlying tool constraints. To this end, we propose tool amplification, a novel paradigm that enhances the collective capabilities of specialized tools through optimized, dynamic coordination within individual tasks. Instantiating this paradigm, we introduce ChemAmp, a computationally lightweight framework that dynamically treats chemistry tools (e.g., UniMol2, Chemformer) as composable building-block agents. It constructs task-specialized super-agents that transcend atomic tool constraints with limited data (≤10 samples). Our evaluations across four core chemistry tasks molecular design, molecule captioning, reaction prediction, and property prediction demonstrate that ChemAmp outperforms chemistry-specialized models, generalist LLMs, and agent systems with tool orchestration. Critically, this bottom-up construction strategy enables 94% inference token cost reductions versus vanilla multi-agent systems.
Reranking is fundamental to information retrieval and retrieval-augmented generation, with recent Large Language Models (LLMs) significantly advancing reranking quality. Most current works rely on large-scale LLMs (>7B parameters), presenting high computational costs. Small Language Models (SLMs) offer a promising alternative because of computational efficiency. However, our preliminary quantitative analysis reveals key limitations of SLMs: their representation space is narrow, leading to reduced expressiveness, and they struggle with understanding task prompts without fine-tuning. To address these issues, we introduce a novel two-stage training approach, ProRank, for SLM-based document reranking. We propose using reinforcement learning to improve the understanding of task prompts. Additionally, we introduce fine-grained score learning to enhance representation expressiveness and further improve document reranking quality. Extensive experiments suggest that ProRank consistently outperforms both the most advanced open-source and proprietary reranking models. Notably, our ProRank even surpasses powerful LLM reranking models on the BEIR benchmark, establishing that properly trained SLMs can achieve superior document reranking performance while maintaining computational efficiency.
Large language models (LLMs) exhibit strong reasoning capabilities but typically require expensive post-training to reach high performance. Recent test-time alignment methods offer a lightweight alternative, but have been explored mainly for preference alignment rather than reasoning. To bridge this gap, we propose Token-level Adaptive Routing (TARo), which steers frozen LLMs toward structured reasoning entirely at inference time. Specifically, we first train reward models on step-wise mathematical traces to capture fine-grained logical consistency signals, then introduce a learnable token-level router that automatically controls the guidance of the reward model to the base model. Extensive experiments show that TARo significantly improves reasoning performance by up to +22.4% over base model and +8.4% over existing token-level test-time alignment methods, while also boosting out-of-distribution clinical reasoning (MedXpertQA) and instruction following (AlpacaEval). Furthermore, TARo also generalizes from small to large backbones without retraining, extending test-time alignment from preference optimization to robust, cross-domain reasoning.
Large language models (LLMs) seem to offer an easy path to interpretability: just ask them to explain their answers. Yet the features driving an answer often differ from those emphasized in its explanation, meaning post-hoc rationales can misrepresent what actually shaped the model’s output. We quantify this gap by comparing the feature-importance distributions of answers and their explanations. Prior analyses reveal such discrepancies, but large-scale study has been limited by the high computational cost of attribution methods. To address this, we introduce the Post-hoc Self-Consistency Bank (PSCB), a large-scale benchmark linking model decisions with diverse explanations and attribution vectors across datasets, methods, and model families. Using PSCB, we find that Spearman rank correlation provides a more reliable signal of alignment than cosine similarity. Building on this insight, we apply Direct Preference Optimization (DPO) to attribution-based preference data, improving alignment without degrading task accuracy, and show that standard supervised fine-tuning on the same data fails to achieve comparable gains. These improvements generalize robustly across domains, paving the way toward scalable and faithful alignment between LLM decisions and their natural language explanations.
Due to the prevalence of large language models (LLMs), key-value (KV) cache reduction for LLM inference has received remarkable attention. Among numerous works that have been proposed in recent years, layer-wise token pruning approaches, which select a subset of tokens at particular layers to retain in KV cache and prune others, are one of the most popular schemes. They primarily adopt a set of pre-defined layers, at which tokens are selected. Such design is inflexible in the sense that the accuracy significantly varies across tasks and deteriorates in harder tasks such as KV retrieval. In this paper, we propose ASL, a training-free method that adaptively chooses the selection layer for KV cache reduction, exploiting the variance of token ranks ordered by attention score. The proposed method balances the performance across different tasks while meeting the user-specified KV budget requirement. ASL operates during the prefilling stage and can be jointly used with existing KV cache reduction methods such as SnapKV to optimize the decoding stage. By evaluations on the InfiniteBench, RULER, and NIAH benchmarks, we show that ASL, equipped with one-shot token selection, adaptively trades inference speed for accuracy, outperforming state-of-the-art layer-wise token pruning methods in difficult tasks.
Do reasoning models have "Aha!" moments?Prior work suggests that models like DeepSeek-R1-Zero undergo sudden mid-trace realizations that lead to accurate outputs, implying an intrinsic capacity for self-correction. Yet, it remains unclear whether such intrinsic shifts in reasoning strategy actually improve performance.Here, we study mid-reasoning shifts and instrument training runs to detect them. Our analysis spans 1M+ reasoning traces, hundreds of training checkpoints, three reasoning domains, and multiple decoding temperatures and model architectures.We find that reasoning shifts are rare, do not become more frequent with training, and seldom improve accuracy, indicating that they do not correspond to prior perceptions of model insight. However, their effect varies with model uncertainty. Building on this finding, we show that artificially triggering extrinsic shifts under high entropy reliably improves accuracy. Our results show that mid-reasoning shifts are symptoms of unstable inference behavior rather than an intrinsic mechanism for self-correction.
Instruction tuning is crucial for optimizing Large Language Models (LLMs), as the quality and diversity of instructional data significantly influence model performance. This naturally underscores the importance of an effective and efficient data selection strategy. However, recent mainstream data selection methods typically rely on LLMs to score instruction quality—taking advantage of their capabilities, but at the cost of high computational overhead and reduced data diversity. To address these limitations, in this paper, we propose MergeIT, a novel LLM-based Merging strategy for better Instruction Tuning that shifts the focus from selection to synthesis. MergeIT consists of two stages: first, topic-aware filtering clusters and refines the dataset, preserving diversity while eliminating redundancy without relying on LLM-based scoring, significantly reducing time and computational cost. Second, LLM-based merging synthesizes semantically similar instructions into more informative and compact training data, enhancing data richness while further reducing the size of the dataset. Experimental results demonstrate that MergeIT enables efficient, diverse, and scalable instruction selection and synthesis, establishing LLM-based merging as a promising alternative to prior scoring-based selection methods for instruction tuning.
High-quality and diverse multimodal data are essential for improving vision–language models (VLMs), yet existing datasets often contain noisy, redundant, and poorly aligned samples. To address these problems, data filtering is commonly used to enhance the efficiency and performance of multimodal learning, but it introduces extra computational cost because filtering models are usually trained on the same data they are meant to screen. To reduce this cost, we study DOSE, which explores whether off-the-shelf pretrained models that have never seen the target data can be used to select training samples for larger and stronger multimodal models without any task-specific training. Even without fine-tuning, these models can effectively assess text quality and image–text alignment to guide data selection. Based on this, we build a joint quality–alignment distribution and apply adaptive weighted sampling to select informative samples while maintaining long-tail diversity. This approach greatly enhances data diversity and enables models trained on DOSE-filtered data to achieve comparable or even better results than those trained on the full dataset in standard VQA and math benchmarks. Extensive experiments demonstrate the effectiveness, efficiency, and scalability of our method.
The ubiquity of text-attributed graph data has highlighted the need for graph learning models with exceptional generalization across diverse textual and structural contexts. Current approaches struggle to extract generalizable insights from heterogeneous graph data, requiring extensive fine-tuning and limiting versatility across domains. In this work, we propose AnyGraph, a unified graph foundation model designed to handle key challenges: i) Structure Heterogenity - addressing distribution shift in graph structural patterns; ii) Feature Heterogenity - handling diverse textual representations; iii) Fast Adaptation - efficiently adapting to new graph-text domains. We build AnyGraph upon a Graph Mixture-of-Experts (MoE) architecture with a lightweight expert routing mechanism that effectively manages cross-domain distribution shift. Extensive experiments on 38 diverse datasets demonstrate AnyGraph’s strong zero-shot performance across domains with significant distribution shift, validating its fast adaptation ability and scaling law emergence. Our model is open-sourced and available at: https://github.com/HKUDS/AnyGraph.
The inherent rigidity of fixed-size benchmarks makes them an inefficient tool for model evaluation. Diverse evaluation objectives, including model ranking, model selection and testing throughout development, demand varying levels of statistical power. The mismatch between fixed sample sizes and these diverse needs results in either excessive computational cost or compromised reliability – a critical concern for model evaluation. To overcome these limitations, we call for adoption of sequential testing in our field. We provide an adaptive evaluation framework, that provides a principled way to navigate the trade-off between efficiency and reliability in model evaluation. Our framework combines the established statistical paradigm of sequential testing with stopping criteria tailored to common evaluation needs such as diminishing returns detection, and minimum detectable effect size. We demonstrate its ability to adaptively manage the efficiency-reliability trade-off on the Open VLM Leaderboard, including, for example, a 80% reduction in computational cost compared to fixed-size evaluation (with a 2.5-point CI width allowance) while maintaining statistical significance.
Bridging the Culture Gap: A Framework for LLM-Driven Socio-Cultural Localization of Math Word Problems in Low-Resource Languages
PDF ↗Large language models (LLMs) have demonstrated significant capabilities in solving mathematical problems expressed in natural language. However, multilingual and culturally-grounded mathematical reasoning in low-resource languages lags behind English due to the scarcity of socio-cultural task datasets that reflect accurate native entities such as person names, organization names, and currencies. Existing multilingual benchmarks are predominantly produced via translation and typically retain English-centric entities, owing to the high cost associated with human annotator-based localization. Moreover, automated localization tools are limited, and hence, truly localized datasets remain scarce. To bridge this gap, we introduce a framework for LLM-driven cultural localization of math word problems that automatically constructs datasets with native names, organizations, and currencies from existing sources. We find that translated benchmarks can obscure true multilingual math ability under appropriate socio-cultural contexts. Through extensive experiments, we also show that our framework can help mitigate English-centric entity bias and improve robustness when native entities are introduced across various languages.
CMTD: Cognitive Modeling with Traits and Distortions for Multimodal Emotion Recognition in Conversations
PDF ↗This paper introduces a new multi-agent framework, CMTD (Cognitive Modeling with Traits and Distortions), for multimodal emotion recognition in conversations (MERC). Instead of relying on shallow analysis of emotions, CMTD reconstructs a cognitive model by taking advantage of stable personality traits, dynamic cognitive distortions, visual and acoustic features of interlocutors to enhance the emotional intelligence of LLMs. CMTD includes trait, distortion detection, vision, and speech agents that provide psychological and multimodal indicators for the fusion agent to make the final prediction. Experimental results on MELD and IEMOCAP show that traits temper negativity bias from distortions, and cognitive modeling with psychological, visual, and acoustic information can improve the performance of MERC.CMTD is flexible and easy to adapt to advanced emotional AI systems (Github link: https://github.com/Shaun-le/CMTD.git).
Autoregressive (AR) language models and Diffusion Language Models (DLMs) constitute the two principal paradigms of large language models. However, both paradigms suffer from insufficient reasoning capabilities. Human reasoning inherently relies on causal knowledge and thought, which are reflected in natural language. But in the AR paradigm, language is modeled as next token prediction (a strictly left-to-right, token-by-token order), whereas natural language itself exhibits more flexible causal structures. In the DLM paradigm, the attention mechanism is fully connected, which entirely disregards causal order. To fill this gap, we propose the Causal Concept-Guided Diffusion Language Model (C^2DLM). Starting from DLM’s fully connected attention, C^2DLM first obtains a concept-level causal graph from the teacher model, and then explicitly guides attention to learn causal relationships between concepts. By focusing on causal relationships and avoiding interference from difficult subgoals involving causal inversion, C^2DLM achieves a 12% improvement and a 3.2× training speedup on the COT-OrderPerturb task, along with an average gain of 1.31% across six downstream reasoning tasks. Code and data are available here.
Understanding In-Context Learning Beyond Transformers: An Investigation of State Space and Hybrid Architectures
PDF ↗We perform in-depth evaluations of in-context learning (ICL) on state-of-the-art transformer, state-space, and hybrid large language models over two categories of knowledge-based ICL tasks. Using a combination of behavioral probing and intervention-based methods, we have discovered that, while LLMs of different architectures can behave similarly in task performance, their internals could remain different. We discover that function vectors (FVs) responsible for ICL are primarily located in the self-attention and Mamba layers, and speculate that Mamba2 uses a different mechanism from FVs to perform ICL. FVs are more important for ICL involving parametric knowledge retrieval, but not for contextual knowledge understanding. Our work contributes to a more nuanced understanding across architectures and task types. Methodologically, our approach also highlights the importance of combining both behavioural and mechanistic analyses to investigate LLM capabilities.
One of the key factors influencing the reasoning capabilities of LLM-based agents is their ability to leverage long-term memory. Integrating long-term memory mechanisms allows agents to make informed decisions grounded in historical interactions. While recent advances have significantly improved the storage and retrieval components—e.g., by encoding memory into dense vectors for similarity search or organizing memory as structured knowledge graphs—most existing approaches fall short in memory updating. In particular, they lack mechanisms for dynamically refining preference memory representations in response to evolving user behaviors and contexts. To address this gap, we propose a Preference-Aware Memory Update Mechanism (PAMU) that enables dynamic and personalized memory refinement. By integrating sliding window averages (SW) with exponential moving averages (EMA), PAMU constructs a fused preference-aware representation that captures both short-term fluctuations and long-term user tendencies. We conduct experiments on five task scenarios of the LoCoMo dataset, and the results show that our mechanism can significantly improve the output quality of LLM in five baselines, validating its effectiveness in long-term conversations.