Despite the advancing reasoning capabilities of large language models (LLMs), many reasoning evaluations are increasingly compromised by data contamination, which induces unreliable on leaked inputs. While this phenomenon is widely observed, its underlying mechanism remains poorly understood, hindering the ability to distinguish generalization from memorization and to develop effective solutions. In this work, we first identify a distinctive signal of contaminated reasoning, namely the mutual information decay between representations and gradients. Our mechanistic analysis reveals that contaminated models exhibit pronounced eigenspectrum concentration in their representations, leading to a low-dimensional computation regime. Under leaked inputs, this mechanism weakens the linear coupling between representations and gradients, manifested as a structural decay of the singular values in the whitened space. We show that this narrowing geometry mathematically implies a reduction in mutual information, and further demonstrate the practical utility of our analysis by successfully restoring the reasoning behavior of contaminated models, achieving an 11.03% improvement in average consistency with the base model over the strongest baseline.

Accurate and effective discrete image tokenization is crucial for long image sequence processing. However, current methods rigidly compress all content at a fixed rate, ignoring the variable information density of images and leading to either redundancy or information loss. Inspired by information entropy, we propose TaTok, a Theoretically grounded adaptive image Tokenization framework. We rigorously identify two key drawbacks in existing methods: information insufficiency when reconstructing images with patch tokens alone, and information redundancy among patch tokens. To address these, we introduce global tokens that model mutual information across patch tokens, and a Dynamic Token Filtering (DTF) algorithm based on cumulative conditional entropy to eliminate redundancy. Experiments confirm TaTok's state-of-the-art performance, delivering a 1.3x gFID improvement and 8.7x inference speedup. By allocating tokens according to information richness, TaTok enables more compressed yet accurate image tokenization, offering valuable insights for future research.

Event Causality Extraction (ECE) aims to extract causal event pairs from text. Existing methods overlook the interplay between causal event pairs and their corresponding textual evidence (e.g., causal event mention pairs), and fail to effectively leverage global causal dependency information. To address these issues, we propose a Mention-Level Causal Evidence and Global Causal Graph Reasoning (MLCE-GCGR) framework to enhance ECE. First, we introduce an auxiliary Event Mention Causality Extraction (EMCE) task, which extracts causal event mention pairs, to provide evidence for the main ECE task, and design a Dual-Level Interaction Enhancement (DLIE) strategy to enhance the bidirectional interplay between event-level and mention-level causality. Second, we develop a Global Causal Graph Reasoning (GCGR) module that simulates human-like multi-turn reasoning, aiming to progressively refine the causal graph by capturing global dependencies among event mentions, types, and arguments. Experiments on four benchmark datasets show that our method outperforms state-of-the-art approaches. Moreover, by extracting causal event mention pairs as supporting evidence, our approach improves the interpretability of structured causality extraction.

Multi-turn long-horizon planning is critical for foundation model agents, yet how to fundamentally improve it remains unclear. Existing models are trained on uncontrollable and opaque Internet data, making it difficult to identify how planning ability is acquired, shaped, and integrated. To address this challenge, we introduce a unified and controlled multi-turn environment that enables precise control. It allows systematically study long-horizon planning across three stages. (1) Planning ability acquisition during pre-training. We study data format, distribution, and quality. Explicit world model construction through CoT state transition modeling yields stronger long-horizon generalization. Atomic skills alone are insufficient for compositional generalization, whereas a litte long-horizon data works. Moreover, suboptimal trajectories severely impair performance because errors amplify over long horizons. (2) Planning ability shaping via GRPO and OPD post-training. Through mutual information, we distinguish general planning patterns from task-specific planning knowledge. For planning patterns, we identify three application regions of post-training: unnecessary, effective, and unsupported. OPD has a broader effective region than GRPO under low-quality and long-horizon settings, as it provides more consistent update directions. For planning knowledge, distilling unseen procedures from a teacher with different knowledge may impair student's prior world modeling without fully establishing new knowledge. (3) Planning ability integration through MOPD post-training. We show that multi-teacher on-policy distillation (MOPD) integrates capabilities by converging to shared planning-pattern across environments. Compatible patterns enable cross-environment generalization, partially shared patterns support continual learning, while completely conflicting patterns cause severe interference.

Understanding why real-world events occur is important for both natural language processing and practical decision-making, yet direct-cause inference remains underexplored in evidence-rich settings. To address this gap, we organized SemEval-2026 Task 12: Abductive Event Reasoning (AER).[The task data is available at https://github.com/sooo66/semeval2026-task12-dataset.git] The task asks systems to identify the most plausible direct cause of a target event from supporting evidence. We formulate AER as an evidence-grounded multiple-choice benchmark that captures key challenges of real-world causal reasoning, including distributed evidence, indirect background factors, and semantically related but non-causal distractors. The shared task attracted 122 participants and received 518 submissions. This paper presents the task formulation, dataset construction pipeline, evaluation setup, and system results. AER provides a focused benchmark for abductive reasoning over real-world events and highlights challenges for future work on causal reasoning and multi-document understanding.

Post-training quantization reduces the deployment cost of large language models, yet how severely a quantized model degrades is not determined by bit-width alone. We systematically study weight-only post-training quantization across bit-widths, quantization methods, model scales and downstream tasks on multiple model families. We observe that such degradation varies substantially across these factors: 4-bit quantization usually preserves performance, 2-bit often causes broad degradation, and at 3-bit, degradation becomes apparent but varies markedly with task type, quantization method and model scale. To explain this variability, we use the signal-to-noise ratio (SNR) to measure how strongly quantization perturbs full-precision representations. We trace degradation back to two linked processes: how quantization errors arise within individual modules, and how they accumulate across layers. First, a source SNR decomposition shows that newly introduced errors depend on three factors: the magnitude of the weight error, the strength of the task-specific signal, and how strongly the quantization error aligns with task-specific activations. Different factors affect these components in distinct ways. Second, a cross-layer propagation analysis shows that these errors can be attenuated, preserved, or amplified as they pass across layers, and that larger models benefit from weaker error amplification. Together, these results establish that quantization degradation is governed by how errors are introduced at the source and how they accumulate across the network.

Legal AI tasks require large language models (LLMs) to demonstrate robust legal reasoning for complex, knowledge-intensive, and multi-task scenarios such as legal judgment prediction (LJP), statute article generation (SAG), and legal document generation (LDG). However, the field’s emphasis on logical consistency, coupled with persistent hallucination issues, limits the practicality of conventional retrieval-augmented generation (RAG), which struggles with lengthy legal texts, limited context windows, and high computational costs. To address these challenges, we propose Legal-AP, a framework that enhances multi-task legal performance and reasoning quality through knowledge Augmentation and adapter-wise Parametric fusion. Instead of appending retrieved documents to the input, Legal-AP injects external legal knowledge and reasoning-critical patterns directly into model parameters via Low-Rank Adaptation (LoRA). Multiple task-specific adapters are dynamically composed to enable efficient, domain-aware reasoning across diverse legal tasks. Furthermore, we construct Legal-CA, a multi-task dataset covering criminal, administrative, and civil domains, with expert annotations emphasizing both substantive legal and reasoning aspects. Experiments show that Legal-AP consistently surpasses conventional RAG approaches and achieves comparable or superior performance to GPT-4o on several tasks, underscoring the promise of parametric knowledge integration for advancing reasoning-driven Legal AI.

Explainable diagnosis requires that authoritative medical knowledge provide the rationales linking a patient’s clinical manifestations to the diagnostic conclusion. Although large language models (LLMs) hold great potential to facilitate explainable diagnosis, their effectiveness is often constrained by insufficient diagnostic expertise. To address this limitation, we propose Self-learned Explainable Knowledge Augmented Diagnosis (SEKAD), a unified LLM-based framework for faithful and explainable diagnosis. Our approach builds a high-quality diagnostic knowledge base through a record-driven explanation learning paradigm, as well as applies this knowledge via an explanation-based diagnostic process that ensures faithful inference. Experiments on the DiReCT and JAMA benchmarks show that SEKAD consistently outperforms strong baselines across the metrics. In particular, on the DiReCT benchmark, SEKAD improves the explanation completeness metric from 64.5% to 76.9% over the best existing methods, highlighting its effectiveness in enhancing diagnostic explainability and showing that our text mining approach produces knowledge that is both reliable in quality and large in quantity.

Retrieval-augmented generation (RAG) plays a critical role in user-generated content (UGC) platforms, but its effectiveness critically depends on accurate query-document relevance assessment. Despite recent advances in applying large language models (LLMs) to relevance modeling, UGC platforms present unique challenges: 1) ambiguous user intent due to sparse user feedback in RAG scenarios, and 2) asymmetric relevance, where relevance is driven by localized answer-bearing content rather than global query-document similarity. To address these issues, we propose the Reinforced Reasoning model for Relevance Assessment (R3A), which decomposes relevance assessment into intent inference and evidence grounding. R3A leverages auxiliary high-clicked documents to infer latent query intent, and extracts verbatim evidence fragments to ground relevance decisions, reducing noise sensitivity and improving asymmetric relevance modeling. Experimental results demonstrate that R3A substantially outperforms strong baselines on offline benchmarks, while the distilled R3A-1.5B model achieves significant gains in large-scale online A/B testing, effectively balancing performance and practical deployability.

Recent works train agents by constructing large-scale multimodal environment pools. However, we find that simply increasing the number of multimodal environments does not always benefit. We further analyze the limitations in current multimodal environment distributions through a series of experiments. Based on these findings, we study how to build more effective training environment distributions from two dimensions: **diversity** and **difficulty structure**. For diversity, we propose **Ability-aware Environment Selection (AES)** to obtain diverse environment sets. For difficulty structure, we propose **Hierarchical Difficulty Curriculum (HDC)**, which organizes curriculum learning through two difficulty levels: harness weakening and state-scale progression. Experiments show that AES and HDC effectively improve multimodal agent training.

Post-Training Quantization (PTQ) is critical for the efficient deployment of Large Language Models (LLMs). While 4-bit quantization is widely regarded as an optimal trade-off, reducing the precision to 2-bit usually triggers a catastrophic “performance cliff.” It remains unclear whether the underlying mechanisms differ fundamentally. Consequently, we conduct a systematic mechanistic analysis, revealing two qualitatively distinct failure modes: Signal Degradation, where the computational patterns remain intact but information precision is impaired by cumulative error; and Computation Collapse, where key components fail to function, preventing correct information processing and destroying the signal in the early layers. Guided by this diagnosis, we conduct mechanism-aware interventions, demonstrating that targeted, training-free repair can mitigate Signal Degradation, but remains ineffective for Computation Collapse. Our findings provide a systematic diagnostic framework for PTQ failures and suggest that addressing Computation Collapse requires structural reconstruction rather than mere compensation.

Recent AI agents can flexibly invoke skills to solve complex tasks, but their long-term improvement is fundamentally constrained by a lack of systematic skill construction, accumulation, and transfer. In particular, without a unified framework for skill consolidation, agents tend to redundantly construct similar capabilities across different tasks, are unable to effectively transform experience into reusable assets, and struggle to generalize task-specific skills to novel scenarios. To address this limitation, we propose SkillPyramid, a skill consolidation framework that reuses existing skill experience for broader task generalization. Operating on a hierarchical skill topology, SkillPyramid further introduces a self-evolution mechanism that enables agents to compose, validate, and incorporate new skills during task execution. Experiments on ALFWorld, WebShop, and ScienceWorld across four backbone models show that SkillPyramid substantially increases the average reward by 38.0

Large Language Model-based agents have demonstrated impressive capabilities across various domains. To further enhance their abilities, the collaboration of multiple agents presents a promising avenue. Recently, Multi-Agent Debate (MAD) was introduced as a representative collaborative paradigm, wherein agents engage in multiple rounds of discussion to explore potential solutions to a problem. However, researchers observed that MAD is not stably superior to single-agent methods. Unfortunately, there has been insufficient investigation of this issue. In light of this limitation, we for the first time empirically identify what leads to the instability of MAD: the woozle effect, which refers to the propagation of hallucinations among agents during debate. Because MAD employs a static and fully connected communication topology, agents are constrained to passively receive information from all connected peers, which increases the risk of being misled by hallucinations and subsequent propagation. To mitigate this challenge, we propose DIGRA, a novel MAD framework with Dynamic communication topology driven by the Information Gain RAtio. Our evaluations across various benchmarks demonstrate that appropriately selecting communication partners significantly mitigates hallucination propagation and yields superior collective intelligence.

Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) exhibit fundamentally different behaviors in enhancing multi-task reasoning for large language models (LLMs). Our preliminary experiments revealed a phenomenon: SFT suffers from severe task conflicts under multi-stage training, whereas RL enables stable coexistence across diverse tasks. Empirically, we trace this to the parameter level, observing that RL induces sparse and approximately orthogonal updates across tasks. We provide a theoretical explanation for this mechanism by analyzing multi-task gradient interference. Our results reveal a distinction: interference in SFT is norm-limited, scaling with the absolute gradient magnitude, whereas interference in RL is variance-limited, bounded by the gradient variance induced by advantage normalization and on-policy optimization. This small variance bound yields near-orthogonal optimization directions across tasks. Leveraging this insight, we propose Parallel-RL, a paradigm that decouples multi-task training, significantly improving efficiency and flexibility.

Parameter-Efficient finetuning (PEFT) enhances model performance on downstream tasks by updating a minimal subset of parameters. Representation finetuning (ReFT) methods further improve efficiency by freezing model weights and optimizing internal representations with fewer parameters than PEFT, outperforming PEFT on several tasks. However, ReFT exhibits a significant performance decline on mathematical reasoning tasks. To address this problem, the paper demonstrates that ReFT's poor performance on mathematical tasks primarily stems from its struggle to generate effective reasoning prefixes during the early inference phase. Moreover, ReFT disturbs the numerical encoding and the error accumulats during the CoT stage. Based on these observations, this paper proposes Bias-REstrained Prefix Representation FineTuning (BREP ReFT), which enhances ReFT's mathematical reasoning capability by truncating training data to optimize the generation of initial reasoning prefixes, intervening on the early inference stage to prevent error accumulation, and constraining the intervention vectors' magnitude to avoid disturbing numerical encoding. Extensive experiments across diverse model architectures demonstrate BREP's superior effectiveness, efficiency, and robust generalization capability, outperforming both standard ReFT and weight-based PEFT methods on the task of mathematical reasoning.

This paper notices that while symbolic instruction and neural parameters play different roles on steering LLMs’ behavior, both instructions and parameters are the compression of task data, they are supposed be strongly correlated and can be learned to predict one from the other. Therefore, This paper proposes a novel neural network framework, SHIP ( Sh uttle between the I nstructions and the P arameters), to model and learn the bi-directional mappings between the instructions and the parameters of LLMs. We verify that SHIP can effectively map one of the instructions/parameters to the other by evaluating it on the tasks of instruction deduction and induction. The results show that SHIP performs better than existing baseline methods in terms of deductive capabilities while significantly surpassing them in inductive capabilities. Moreover, SHIP can effectively combine the two mapping processes to perform excellent inductive reasoning. We further discuss how the latent fusing methods and latent dimensions affect SHIP’s performance, and show SHIP can effectively generalize with pre-training. The code and data for this paper are released at https://anonymous.4open.science/r/Shuttle-Between-Instructions-Parameters

Diffusion large language models (dLLMs) are a promising alternative to autoregressive generation. However, reasoning-oriented post-training for dLLMs remains challenging. Supervised fine-tuning (SFT) for dLLMs requires dense but often off-policy masked states, while reinforcement learning (RL) relies on sparse rewards or value modeling. This paper proposes trace-based on-policy distillation (TOPD), a teacher-supervised framework that transfers reasoning ability to a target dLLM without reward estimation. The key idea is to supervise a dLLM on its own denoising trajectory, focusing on the trace-aligned token decisions that form the final response. Specifically, TOPD samples on-policy diffusion trajectories from the target dLLM, obtains teacher token distributions from a teacher model on the corresponding partially denoised states, and updates the target dLLM with a token-level Reverse Kullback-Leibler (Reverse-KL) objective. This design preserves dense teacher supervision while aligning training with the model's own denoising states. On mathematical reasoning benchmarks, TOPD enables SDAR-4B-Chat to match the MATH500 accuracy of its RL-trained counterpart TraDo-4B-Instruct, with gains of +5.7 under static evaluation and +4.5 under dynamic evaluation. Compared with the RL-trained counterpart, TOPD achieves this with 4× fewer rollout rounds, corresponding to an estimated 96.0× to-accuracy model-compute speedup.

Large language models (LLMs) excel at text understanding and generation, yet still struggle to reliably understand and apply externally provided procedural rules at scale. To evaluate this capability, we introduce RuleWorld, a large-scale benchmark that reformulates rules as globally reusable abstract units rather than instance-specific facts. In RuleWorld, several scenarios, including single-rule, parallel multi-rule, and multi-hop reasoning, are settled for comprehensive evaluation. We further propose DynaRule, an end-to-end framework that injects the given rules into the KV cache and turns retrieval into an internal, learnable, step-wise process. Specifically, DynaRule employs Stacked Step-Level Attention Training with a special token to enable dynamic rule re-attention and updating during inference. In this way, the model can re-attend to the most relevant rules at each step, dynamically replacing outdated ones to support more stable multi-step reasoning. Experiments on RuleWorld show that existing LLMs face challenges under large rule pools, while DynaRule improves average QA accuracy by up to 19 points and achieves over 85

Synthesizing long-context supervised fine-tuning (SFT) data is a scalable way to enhance the long-context understanding of large language models (LLMs), yet existing approaches share three limitations: narrow task coverage, insufficient instruction difficulty, and a lack of faithfulness supervision. We propose LongCrafter, a structured synthesis framework that couples a hierarchical task taxonomy with an evidence-grounded pipeline. The taxonomy organizes long-context understanding into local/shallow and global/deep levels and yields 32 fine-grained task types that serve as a global generative prior. Guided by this taxonomy, LongCrafter constructs task-aligned long contexts, decomposes them into explicit evidence graphs that model cross-paragraph dependencies, and generates instruction–response pairs strictly grounded in the located evidence spans, ensuring both controllable difficulty and faithful, traceable reasoning. Models fine-tuned on LongCrafter data outperform all SFT baselines and even the official post-trained models on LongBench, LongBench v2, and LooGLE across both Qwen2.5-7B and LLaMA-3.1-8B, with the largest gains on high-difficulty tasks. Further analysis shows that LongCrafter data is more diverse and better spread across difficulty levels, and that the trained models locate evidence robustly regardless of position, effectively mitigating the “lost in the middle” problem.