Legal judgment prediction (LJP) is important for supporting judicial decision-making and improving judicial efficiency. Existing LLM-based approaches often rely on manually specified reasoning pipelines or external legal structures, while feedback-based methods typically provide generic signals of correctness or coherence rather than case-specific legal supervision. We propose IALR: Instance-Adaptive Legal Rubric Supervision for Legal Judgment Prediction. For each training instance, IALR contrasts model responses of varying quality and distills their legally consequential failures into a weighted Instance-Adaptive Legal Rubric. These rubrics convert judgment-relevant conditions into discriminative preference signals for Direct Preference Optimization, enabling the model to learn charge prediction without following a predefined reasoning paradigm. At inference time, IALR requires no additional evaluator or correction module. Experiments on four Chinese criminal charge-prediction benchmarks show that IALR consistently improves Qwen2.5-7B, yielding average absolute accuracy gains of 5.68% in-domain and 8.73% out-of-domain. These results demonstrate the effectiveness and cross-domain transferability of Instance-Adaptive Legal Rubric supervision. The code and data are available at https://github.com/Baii668/IALR.git.

With the evolution of large language models (LLMs), their robustness against individual simple biases has been enhanced. However, we observe that the ensemble of multiple simple biases still exerts a significant adverse impact on LLMs. Given that real-world data samples are typically confounded by a wide range of biases, LLMs tend to exhibit unstable performance when deployed in high-stakes real-world scenarios such as clinical diagnosis and legal document analysis. However, previous benchmarks are constrained to datasets where each sample is manually injected with only one type of bias. To bridge this gap, we propose a multi-bias benchmark where each sample contains multiple types of biases. Experimental results reveal that existing LLMs and debiasing methods perform poorly on this benchmark, highlighting the challenge of eliminating such compounded biases.

Existing causal datasets primarily focus on the commonsense domain, where the questions mainly involve simple, single-hop direct causal relationships. When models possess the corresponding knowledge, even if they cannot understand the causal relationships, they can directly arrive at the correct answers through knowledge matching. However, LLMs often perform poorly when answering questions with complex causal structures and domain-specific expertise. To address the above challenges, we propose MDC-Bench, a multidisciplinary causal evaluation benchmark. MDC-Bench adopts a three-level causal framework consisting of 4 core causal tasks, while its sample content covers 7 representative disciplines and diverse causal structures. In view of the limited coverage of multidisciplinary knowledge during the pre-training phase, the model cannot answer questions relying on knowledge matching. The diverse causal structures force the models to grasp the internal causal logic. We also increase the task complexity through methods such as compound causal operations, aiming to enhance the discriminability among models. MDC-Bench achieves the improvement in terms of domain specialization, structural diversity, and task complexity. Through extensive evaluation, we observe that even the advanced models have substantial room for improvement. MDC-Bench not only establishes a standardized baseline for causal research but also provides valuable insights for the applying LLMs in multiple domains.

While reasoning-augmented large language models (RLLMs) significantly enhance complex task performance through extended reasoning chains, they inevitably introduce substantial unnecessary token consumption—particularly for simpler problems where Short Chain-of-Thought (Short CoT) suffices. This ”overthinking” phenomenon leads to inefficient resource usage without proportional accuracy gains. To address this issue, we propose Self-Route, a dynamic reasoning framework that automatically selects between general and reasoning modes based on model capability estimation. Our approach introduces a lightweight pre-inference stage to extract capability-aware embeddings from hidden layer representations, enabling real-time evaluation of the model’s ability to solve problems. We further construct Gradient-10K, a model difficulty estimation-based dataset with dense complexity sampling, to train the router for precise capability boundary detection. Extensive experiments demonstrate that Self-Route achieves comparable accuracy to reasoning models while reducing token consumption by 30-55% across diverse benchmarks. The proposed framework demonstrates consistent effectiveness across models with different parameter scales and reasoning paradigms, highlighting its general applicability and practical value.

Multi-agent debate systems are used to competitively discuss and resolve complex tasks. However, existing approaches often exhibit pseudo-debate behavior: agents either ignore opponents’ arguments or compromise too readily without critical evaluation. To address this, we propose a three-stage debate framework grounded in abstract argumentation theory. First, Debate Structuring uses principled prompts to elicit well-formed debate interactions. Second, Debate Imitation filters high-quality debates to fine-tune LLMs on proper argumentative conduct. Third, Debate Internalization employs a debate-aware reinforcement learning objective to instill adversarial reasoning patterns. To better quantify debate capability, we introduce two novel metrics: Rebuttal Engagement Score (RES), assessing how thoroughly models engage with opposing arguments, and Belief Governance Score (BGS), evaluating whether models appropriately retain or revise beliefs based on argument strength. Extensive evaluations across six benchmarks show that our framework not only improves task performance but also cultivates more authentic and robust debate behavior in LLMs. Our work contributes a theory-inspired training framework and evaluation suite for advancing multi-agent debate systems.

Multi-agent LLM systems delegate through natural language, creating trust boundaries that single-agent safety hierarchies were not designed to preserve. We identify provenance erasure as a multi-agent safety failure mechanism: even in the minimal two-agent case, delegation strips trust metadata—98% of format markers, 100% of authority entities, and 44% of source attribution across 460 pairs—while often preserving the actionable payload. This lossy transform enables Trust Cascade, amplifying system-channel attacks from 0% to 33–63% cascade ASR across eight open-weight models; a delegation-as-user control isolates channel trust escalation rather than content reformulation. Because the failure occurs at the delegation boundary, we propose the Trust Boundary Auditor (TBA), a zero-inference provenance diagnostic achieving 86–100% conditional detection with 0% false positives across four domains. Training, scale, and closed-source evaluations show that stronger models can interrupt different cascade stages but do not remove the boundary problem, suggesting that multi-agent safety requires delegation gating plus provenance auditing.

Instruction Following (IF) is a core capability of LLMs, requiring strict adherence to diverse constraints, ranging from verifiable ones (e.g., output length) to unverifiable ones (e.g., tone). Reinforcement learning with verifiable rewards has emerged as a paradigm for IF tasks, leveraging LLM-as-a-judge to assess unverifiable constraints. However, we empirically find that this approach remains a significant bottleneck, suffering from severe reward hacking and higher computational overhead. In this work, we first analyze the generalization capabilities of unverifiable constraints and discover that specific constraints exhibit distinct, high-generalization patterns. Motivated by this, we propose TinyJudge, a framework that employs an ensemble of specialized tiny language models (e.g., 0.6B) to provide rewards for soft constraints. By distilling expertise from frontier models into these tiny models, it achieves high-precision, lightweight evaluation. Extensive evaluations across five benchmarks demonstrate that TinyJudge outperforms the baselines by ~10% in average performance and 12% in reward precision. Crucially, it also achieves a 3× speedup in total training time. Our work provides a scalable and robust path for aligning LLMs with unverifiable human instructions.

While Hybrid Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL) has become the standard paradigm for training LLM agents, effective mechanisms for data allocation between these stages remain largely underexplored. Current data arbitration strategies often rely on surface-level heuristics that fail to diagnose intrinsic learning needs. Since SFT targets pattern consolidation through imitation while RL drives structural adaptation via exploration, misaligning data with these functional roles causes severe optimization interference. We propose PRISM, a dynamics-aware framework grounded in Schema Theory that arbitrates data based on its degree of cognitive conflict with the model’s existing knowledge. By analyzing the spatial geometric structure of gradients, PRISM identifies data triggering high spatial concentration as high-conflict signals that require RL for structural restructuring. In contrast, data yielding diffuse updates is routed to SFT for efficient consolidation. Extensive experiments on WebShop and ALFWorld demonstrate that PRISM achieves a Pareto improvement, outperforming state-of-the-art hybrid methods while reducing computational costs by up to 3.22 ×. Our findings suggest that disentangling data based on internal optimization regimes is crucial for scalable and robust agent alignment.

Group-Relative Policy Optimization (GRPO) has emerged as an efficient paradigm for aligning Large Language Models (LLMs), yet its efficacy is primarily confined to domains with verifiable ground truths. Extending GRPO to open-domain settings remains a critical challenge, as unconstrained generation entails multi-faceted and often conflicting objectives—such as creativity versus factuality—where rigid, static reward scalarization is inherently suboptimal. To address this, we propose MAESTRO (Meta-learning Adaptive Estimation of Scalarization Trade-offs for Reward Optimization), which introduces a meta-cognitive orchestration layer that treats reward scalarization as a dynamic latent policy, leveraging the model’s terminal hidden states as a semantic bottleneck to perceive task-specific priorities. We formulate this as a contextual bandit problem within a bi-level optimization framework, where a lightweight Conductor network co-evolves with the policy by utilizing group-relative advantages as a meta-reward signal. Across seven benchmarks, MAESTRO consistently outperforms single-reward and static multi-objective baselines, while preserving the efficiency advantages of GRPO, and in some settings even reducing redundant generation.

In this paper, we explore using Large Language Models (LLMs) for rumor detection on social media. It involves assessing the veracity of claims on social media based on social context (e.g., comments, propagation patterns). LLMs, despite their impressive capabilities in text-based reasoning tasks, struggle to achieve promising rumor detection performance when facing long structured social contexts. Our preliminary analysis shows that large-scale contexts hinder LLMs’ reasoning abilities, while moderate contexts perform better for LLMs, highlighting the need for refined contexts. Accordingly, we propose a semantic-propagation collaboration-base framework that integrates small language models (e.g., graph attention network) with LLMs for effective rumor detection. It models contexts by enabling text semantic and propagation patterns to collaborate through graph attention mechanisms, and reconstruct the context by aggregating attention values during inference. Also, a cluster-based unsupervised method to refine context is proposed for generalization. Extensive experiments demonstrate the effectiveness of proposed methods in rumor detection. This work bridges the gap for LLMs in facing long, structured data and offers a novel solution for rumor detection on social media.

Recent studies provide large language models (LLMs) with textual task-solving experiences via prompts to improve their performance.However, previous methods rely on substantial human labor or time to gather such experiences for each task, which is impractical given the growing variety of task types in user queries to LLMs.To address this issue, we design an autonomous experience transfer framework to explore whether LLMs can mimic human cognitive intelligence to autonomously transfer experience from existing source tasks to newly encountered target tasks. This not only allows the acquisition of experience without extensive costs of previous methods, but also offers a novel path for the generalization of LLMs.Experimental results on 13 datasets demonstrate that our framework effectively improves the performance of LLMs. Furthermore, we provide a detailed analysis of each module in the framework.

Large language models (LLMs) have achieved significant performance in various natural language reasoning tasks. However, they still struggle with performing first-order logic reasoning over formal logical theories expressed in natural language. This is because the previous LLMs-based reasoning systems have the theoretical incompleteness issue. As a result, it can only address a limited set of simple reasoning problems, which significantly decreases their generalization ability. To address this issue, we propose a novel framework, named Generalizable and Faithful Reasoner (GFaiR), which introduces the paradigm of resolution refutation. Resolution refutation has the capability to solve all first-order logic reasoning problems by extending reasoning rules and employing the principle of proof by contradiction, so our system’s completeness can be improved by introducing resolution refutation. Experimental results demonstrate that our system outperforms previous works by achieving state-of-the-art performances in complex scenarios while maintaining performances in simple scenarios. Besides, we observe that GFaiR is faithful to its reasoning process.