While Large Language Models (LLMs) have demonstrated impressive general capabilities, their direct application in the legal domain is often hindered by a lack of precise domain knowledge and complexity of performing rigorous multi-step judicial reasoning. To address this gap, we present LegalOne, a family of foundational models specifically tailored for the Chinese legal domain. LegalOne is developed through a comprehensive three-phase pipeline designed to master legal reasoning. First, during mid-training phase, we propose Plasticity-Adjusted Sampling (PAS) to address the challenge of domain adaptation. This perplexity-based scheduler strikes a balance between the acquisition of new knowledge and the retention of original capabilities, effectively establishing a robust legal foundation. Second, during supervised fine-tuning, we employ Legal Agentic CoT Distillation (LEAD) to distill explicit reasoning from raw legal texts. Unlike naive distillation, LEAD utilizes an agentic workflow to convert complex judicial processes into structured reasoning trajectories, thereby enforcing factual grounding and logical rigor. Finally, we implement a Curriculum Reinforcement Learning (RL) strategy. Through a progressive reinforcement process spanning memorization, understanding, and reasoning, LegalOne evolves from simple pattern matching to autonomous and reliable legal reasoning. Experimental results demonstrate that LegalOne achieves state-of-the-art performance across a wide range of legal tasks, surpassing general-purpose LLMs with vastly larger parameter counts through enhanced knowledge density and efficiency.

Agentic AI systems powered by large language models are increasingly expected to solve capability-intensive tasks by using reusable external skills, which package instructions, tool-use procedures, executable code, and auxiliary resources. However, mainstream skill-based systems typically assume that available skills can be directly placed in the prompt. As skill libraries grow, this assumption becomes impractical: large corpora cannot be fully exposed to the model, while excessive candidate exposure can distract the agent and degrade skill selection. Therefore, scalable skill augmentation requires retrieving a small set of task-relevant skills from a large corpus before incorporating them into the agent’s context. We formulate this paradigm as Skill Retrieval Augmentation (SRA), where agents dynamically retrieve, incorporate, and apply reusable skills from large external corpora on demand. To systematically study SRA, we introduce SRA-Bench, which enables decomposed evaluation across three stages: skill retrieval, skill incorporation, and skill application. SRA-Bench contains 5,400 capability-intensive instances from six domains, 636 manually constructed gold skills, and a corpus of 26,262 skills. Experiments across multiple LLM and skill-use strategies show that SRA can consistently improve agent performance, but current agents struggle to realize this promise in practical pipelines. Specifically, they remain sensitive to distractor skills, lack need-aware skill loading, and often fail to translate relevant skills into correct task execution. Together, these results establish SRA as a distinct research problem and provide a foundation for developing scalable skill-augmented agents.

Large Language Models (LLMs) are exhibiting emergent human-like abilities and are envisioned as the tool for simulating an individual’s communication patterns, behaviors, and personality traits. However, current evaluations of LLM-based persona simulation remain limited: most rely on synthetic dialogues and lack fine-grained analysis of the capability for persona simulation. To address these limitations, we introduce TwinVoice, a comprehensive benchmark for assessing persona simulation across diverse real-world contexts. TwinVoice encompasses three dimensions: Social Persona (public social interactions), Interpersonal Persona (private dialogues), and Narrative Persona (role-based expression). It further decomposes the evaluation into six fundamental capabilities, including opinion consistency, memory recall, logical reasoning, lexical fidelity, persona tone, and syntactic style. Experimental results reveal that while advanced models achieve moderate accuracy in persona simulation, they still fall short of capabilities such as syntactic style and memory recall. Our data, code, and evaluation results are available.

Mock trial has long served as an important platform for professional legal training and education. Traditional mock trials are difficult to access by the public because they rely on professional tutors and human participants. Fortunately, the rise of large language models (LLMs) provides new opportunities for creating more accessible and scalable court simulations. While promising, existing research ignored the systematic design and procedure evaluation of court simulations, which are critical to the credibility and usage of court simulation in practice. To this end, we propose a novel court simulation paradigm, i.e. SimCourt, based on the real-world procedure structure of Chinese courts, and design a comprehensive evaluation framework focusing on both legal judgment prediction and court procedure analysis. Experiments show that our framework can generate simulated trials that better guide the system in predicting the imprisonment, probation, and fine of each case. Further procedure evaluations show that agents’ responses under our simulation framework even outperform judges and lawyers from the real trials in many aspects. These demonstrate the potential of LLM-based court simulation.

As large language models (LLMs) are increasingly applied to real-world legal tasks, evaluating the reliability of their open-ended legal responses has become essential. These tasks require context-sensitive answers and allow little room for error, motivating fine-grained and diagnostic evaluation that can identify specific sources of response quality failures. We introduce LexRubric, a rubric-based benchmark for evaluating open-ended Chinese legal tasks. LexRubric contains 649 instances from legal consultation and judicial examination, which reflect both everyday legal needs and professional legal reasoning and cover 14 legal scenarios. It further includes 12,337 expert-written atomic scoring criteria organized under a unified six-dimensional framework, enabling accurate evaluation and diagnostic analysis across tasks and evaluation dimensions. To validate the reliability of the evaluation, we test multiple judge models and compare model-based judgments with human judgments. We further evaluate 18 recent general and legal-domain LLMs on LexRubric. Results show that different models exhibit distinct capability profiles, and that open-ended legal tasks remain challenging for current LLMs. Data is available at: <https://github.com/foggpoy/LexRubric>.

Legal judgment prediction (LJP) aims to predict judicial outcomes from case facts and typically includes law article, charge, and sentencing prediction. While recent methods perform well on the first two subtasks, legal sentencing prediction (LSP) remains difficult due to its need for fine-grained objective knowledge and flexible subjective reasoning. To address these limitations, we propose MSR², a framework that integrates multi-source retrieval and reasoning in LLMs with reinforcement learning. MSR² enables LLMs to perform multi-source retrieval based on reasoning needs and applies a process-level reward to guide intermediate subjective reasoning steps. Experiments on two real-world datasets show that MSR² improves both accuracy and interpretability in LSP, providing a promising step toward practical legal AI. Our code is available at https://github.com/cjj826/MSR2.

As large language models (LLMs) are increasingly used for long-form generation, reliably evaluating long-form outputs has become a critical challenge. LLM-as-a-judge offers a scalable alternative to human evaluation, yet its reliability in long-form output evaluation remains underexamined: existing meta-evaluation benchmarks focus mainly on short-form outputs. Compared with short-form evaluation, long-form evaluation is not merely a matter of output length; it often requires judges to handle more complex document-level demands. In this work, we introduce LongJudgeBench, a comprehensive benchmark for evaluating LLM judges on long-form outputs across diverse real-world scenarios and judging protocols. We systematically evaluate a broad range of LLM judges, covering multiple base models and judging settings. Our results reveal a substantial reliability gap: current LLM judges remain unstable across scenarios, and rubrics or references are helpful but not always sufficient. We hope LongJudgeBench will support future research on more robust, context-aware, and human-aligned LLM-as-a-judge methods. Our code is available at https://github.com/cjj826/LongJudgeBench.

Integrating external tools enables Large Language Models (LLMs) to interact with real-world environments and solve complex tasks. Given the growing scale of available tools, effective tool retrieval is essential to mitigate constraints of LLMs’ context windows and ensure computational efficiency. Existing approaches typically treat tool retrieval as a traditional ad-hoc retrieval task, matching user queries against the entire raw tool documentation. In this paper, we identify three fundamental challenges that limit the effectiveness of this paradigm: (i) the incompleteness and structural inconsistency of tool documentation; (ii) the significant semantic and granular mismatch between user queries and technical tool documents; and, most importantly, (iii) the multi-aspect nature of tool utility, that involves distinct dimensions, such as functionality, input constraints, and output formats, varying in format and importance. To address these challenges, we introduce Multi-Field Tool Retrieval, a framework designed to align user intent with tool representations through fine-grained, multi-field modeling. Experimental results show that our framework achieves SOTA performance on five datasets and a mixed benchmark, exhibiting superior generalizability and robustness. We have open-sourced all the code and prompts in https://github.com/LittleDinoC/MFTR.

Reinforcement Learning with Verifiable Rewards (RLVR) is widely used to improve reasoning across domains, but outcome-only scalar rewards are often sparse and uninformative, especially for failed samples, where they indicate only failure without explaining why the reasoning breaks down. In this paper, we leverage richer verbal feedback to guide RLVR on failed samples and convert feedback-induced progress into trainable learning signals. We propose MulFeRL (Multi-turn Feedback-guided Reinforcement Learning), a multi-turn, event-triggered RLVR framework that combines progress induction for feedback-guided regeneration of failed samples, progress credit assignment for learning from verifier-confirmed progress, and structured feedback injection for integrating feedback into the model’s reasoning process. Trained on sampled OpenR1-Math, MulFeRL outperforms supervised, self-distillation-based, and RLVR baselines in-domain and shows strong out-of-domain generalization. [We release our code at <https://anonymous.4open.science/r/MulFeRL>.]

Continual post-training enables models to absorb emerging knowledge after deployment, but repeatedly updating shared parameters inevitably accumulates weight drift, which causes catastrophic forgetting and degrades general capabilities. Retrieval-augmented generation avoids such parameter drift, yet often lacks the depth of parametric knowledge integration. In this paper, we propose ReGrad (Retrievable Gradients), a new paradigm that treats gradients as retrievable units of knowledge. ReGrad pre-computes document-specific gradients offline, stores them in an indexed Gradient Bank, and retrieves only query-relevant gradients at inference time for temporary weight adaptation. However, raw language-modeling gradients are optimized for token-level document reconstruction rather than for query-driven knowledge use. We therefore introduce a bi-level meta-learning objective that reshapes document-derived gradients into generalizable adaptation signals for downstream tasks. Experiments across general and domain-specific settings show that ReGrad outperforms CPT and RAG baselines, enabling scalable and reversible parametric knowledge injection without accumulating weight drift.

Large language models (LLMs) are largely static and often redo reasoning or repeat mistakes. Prior experience reuse typically relies on external retrieval, which is similarity-based, can introduce noise, and adds latency. We introduce SEAM (Structured Experience Adapter Module), a lightweight, executor-specific plug-in that stores experience in its parameters and generates a structured, instance-tailored experience entry in a single forward pass to guide a frozen LLM executor. SEAM is trained for utility via executor rollouts and GRPO while keeping the executor frozen, and can be further improved with logged-success SFT after deployment. Experiments on mathematical reasoning benchmarks show consistent accuracy gains across executors with low overhead. Extensive ablation and analysis further elucidate the mechanisms underlying SEAM’s effectiveness and robustness.

Knowledge editing enables efficient updates to Large Language Models (LLMs) by modifying specific knowledge without full-model retraining. Among knowledge editing approaches, in-context editing (ICE) stands out for its ability to inject knowledge without modifying the model’s parameters. However, existing ICE approaches directly edit model context without isolating target knowledge from the reasoning path of model inference, resulting in unreliable and low-quality outputs, particularly in multi-hop tasks. To investigate this issue, we analyze the interaction between reasoning path planning and knowledge injection, showing that the reasoning ability of a LLM is usually coupled with its original knowledge, and directly replacing old knowledge with new one could simultaneously hurt the LLM’s performance in task reasoning. Based on these findings, we propose DecKER, a novel ICE framework that separates model reasoning from knowledge editing. Extensive experiments show that DecKER significantly improves multi-hop reasoning performance by mitigating knowledge conflicts and preserving reasoning integrity.

As Large Language Models (LLMs) demonstrate increasingly strong human-like capabilities, the need to align them with human values has become significant. Recent advanced techniques, such as prompt learning and reinforcement learning, are being employed to bring LLMs closer to aligning with human values. While these techniques address broad ethical and helpfulness concerns, they rarely consider simulating individualized human values. To bridge this gap, we propose SimVBG, a framework that simulates individual values based on individual backstories that reflect their past experience and demographic information. SimVBG transforms structured data on an individual to a backstory and utilizes a multi-module architecture inspired by the Cognitive–Affective Personality System to simulate individual value based on the backstories. We test SimVBG on a self-constructed benchmark derived from the World Values Survey and show that SimVBG improves top-1 accuracy by more than 10% over the retrieval-augmented generation method. Further analysis shows that performance increases as additional interaction user history becomes available, indicating that the model can refine its persona over time. Code, dataset, and complete experimental results are available at https://github.com/bangdedadi/SimVBG.

Existing retrieval-augmented code generation (RACG) methods typically use an external retrieval module to fetch semantically similar code snippets used for generating subsequent fragments. However, even for consecutive code fragments, the content often diverges due to logical progression, resulting in a content gap. This gap undermines the performance of current RACG methods, as external retrieval modules based on content matching fail to infer the specific information need of LLMs to generate the next code fragment. Therefore, we propose SelfRACG, a novel paradigm that enables large language models (LLMs) to Self-express their information needs to enhance RACG. Specifically, SelfRACG includes an information need expression module and a two-stage information need-guided training strategy, which encourages LLMs to express their information need. Extensive experiments demonstrate that SelfRACG can retrieve external knowledge that better aligns with the LLM’s own information needs, resulting in superior generation performance compared to vanilla RACG. Moreover, both the training and deployment costs for retrieval in our framework are much lower than those of the strongest retrieval model.

Knowledge Editing is a technique that updates large language models (LLMs) with new information to maintain their world knowledge. This approach avoids the need to rebuild the model from scratch, thereby addressing the high costs associated with frequent retraining. Among these, the in-context editing paradigm stands out for its effectiveness in integrating new knowledge while preserving the model’s original capabilities. Despite its potential, existing in-context knowledge editing methods are often task-specific, focusing primarily on multi-hop QA tasks using structured knowledge triples. Moreover, their reliance on few-shot prompting for task decomposition makes them unstable and less effective in generalizing across diverse tasks. In response to these limitations, we propose EditCoT, a novel knowledge editing framework that flexibly and efficiently updates LLMs across various tasks without retraining. EditCoT works by generating a chain-of-thought (CoT) for a given input and then iteratively refining this CoT process using a CoT editor based on updated knowledge. We evaluate EditCoT across a diverse range of benchmarks, covering multiple languages and tasks. The results demonstrate that our approach achieves state-of-the-art performance while offering superior generalization, effectiveness, and stability compared to existing methods, marking a significant advancement in the field of knowledge updating.

Multi-agent techniques such as role playing or multi-turn debates have been shown to be effective in improving the performance of large language models (LLMs) in downstream tasks. Despite their differences in workflows, existing multi-agent systems constructed from a single base LLM mostly use natural language for agent communication.While this is appealing for its simplicity and interpretability, it also introduces inevitable information loss as one model must down sample its continuous state vectors to discrete tokens before transferring them to the other model.Such losses are particularly significant when the information to transfer is not simple facts, but reasoning logics or abstractive thoughts.To tackle this problem, we propose a new communication protocol that transfers both natural language tokens and token-wise state transition trajectory from one agent to another.Particularly, compared to the actual state value, we find that the sequence of state changes in LLMs after generating each token can better reflect the information hidden behind the inference process.We propose a State Delta Encoding (SDE) method to represent state transition trajectories.The experimental results show that multi-agent systems with SDE achieve SOTA performance compared to other communication protocols, particularly in tasks that involve complex reasoning. We have open-sourced all the code and data in https://github.com/LittleDinoC/StateDelta/.

The use of large language models (LLMs) as automated evaluation tools to assess the quality of generated natural language, known as ”LLMs-as-Judges”, has demonstrated promising capabilities and is rapidly gaining widespread attention. However, when applied to pairwise comparisons of candidate responses, LLM-based evaluators often exhibit selection bias. Specifically, their judgments may become inconsistent when the option positions or ID tokens are swapped, compromising the effectiveness and fairness of the evaluation result. To address this challenge, we introduce CalibraEval, a novel label-free method for mitigating selection bias during inference. Specifically, CalibraEval reformulates debiasing as an optimization task aimed at adjusting observed prediction distributions to align with unbiased prediction distributions. To solve this optimization problem, we propose a non-parametric order-preserving algorithm (NOA). This algorithm leverages the partial order relationships between model prediction distributions, thereby eliminating the need for explicit labels and precise mathematical function modeling. Empirical evaluations of LLMs in multiple representative benchmarks demonstrate that CalibraEval effectively mitigates selection bias and improves performance compared to existing debiasing methods. This work marks a step toward building more robust and unbiased automated evaluation frameworks, paving the way for improved reliability in AI-driven assessments. The code can be found at https://github.com/CSHaitao/CalibraEval.

With the increasing intelligence and autonomy of LLM Agents, their potential applications in the legal domain are becoming increasingly apparent. However, existing general-domain benchmarks are unable to fully capture the complexity and subtle nuances inherent in real-world judicial cognition and decision-making. Therefore, we propose LegalAgentBench, a comprehensive benchmark specifically designed to evaluate LLM Agents in the Chinese legal domain. LegalAgentBench includes 17 corpora from real-world legal scenarios and provides 37 tools for interacting with external knowledge. To cover tasks of varying difficulty and types, we designed a scalable task construction process that enables a more precise evaluation of performance in both tool utilization and reasoning. Moreover, Beyond assessing performance through the success rate of final outcomes, LegalAgentBench incorporates keyword analysis during intermediate processes to calculate progress rates, facilitating a more fine-grained evaluation. We evaluated eight popular LLMs, highlighting the strengths, limitations, and potential areas for improvement of existing models and methods. LegalAgentBench sets a new benchmark for the practical application of LLMs in the legal domain, with its code and data available at https://github.com/CSHaitao/LegalAgentBench.

Statute retrieval aims to find relevant statutory articles for specific queries. This process is the basis of a wide range of legal applications such as legal advice, automated judicial decisions, legal document drafting, etc. Existing statute retrieval benchmarks emphasize formal and professional queries from sources like bar exams and legal case documents, thereby neglecting non-professional queries from the general public, which often lack precise legal terminology and references. To address this gap, we introduce the STAtute Retrieval Dataset (STARD), a Chinese dataset comprising 1,543 query cases collected from real-world legal consultations and 55,348 candidate statutory articles. Unlike existing statute retrieval datasets, which primarily focus on professional legal queries, STARD captures the complexity and diversity of real queries from the general public. Through a comprehensive evaluation of various retrieval baselines, we reveal that existing retrieval approaches all fall short of these real queries issued by non-professional users. The best method only achieves a Recall@100 of 0.907, suggesting the necessity for further exploration and additional research in this area.

Hallucinations in large language models (LLMs) refer to the phenomenon of LLMs producing responses that are coherent yet factually inaccurate. This issue undermines the effectiveness of LLMs in practical applications, necessitating research into detecting and mitigating hallucinations of LLMs. Previous studies have mainly concentrated on post-processing techniques for hallucination detection, which tend to be computationally intensive and limited in effectiveness due to their separation from the LLM’s inference process. To overcome these limitations, we introduce MIND, an unsupervised training framework that leverages the internal states of LLMs for real-time hallucination detection without requiring manual annotations. Additionally, we present HELM, a new benchmark for evaluating hallucination detection across multiple LLMs, featuring diverse LLM outputs and the internal states of LLMs during their inference process. Our experiments demonstrate that MIND outperforms existing state-of-the-art methods in hallucination detection.

Dynamic retrieval augmented generation (RAG) paradigm actively decides when and what to retrieve during the text generation process of Large Language Models (LLMs).There are two key elements of this paradigm: identifying the optimal moment to activate the retrieval module (deciding when to retrieve) and crafting the appropriate query once retrieval is triggered (determining what to retrieve).However, current dynamic RAG methods fall short in both aspects. Firstly, the strategies for deciding when to retrieve often rely on static rules. Moreover, the strategies for deciding what to retrieve typically limit themselves to the LLM’s most recent sentence or the last few tokens, while the LLM’s information needs may span across the entire context.To overcome these limitations, we introduce a new framework, DRAGIN, i.e., Dynamic Retrieval Augmented Generation based on the Information Needs of LLMs. Our framework is specifically designed to make decisions on when and what to retrieve based on the LLM’s information needs during the text generation process.We evaluate DRAGIN along with existing methods comprehensively over 4 knowledge-intensive generation datasets. Experimental results show that DRAGIN achieves superior performance on all tasks, demonstrating the effectiveness of our method.

For text-to-image generation, automatically refining user-provided natural language prompts into the keyword-enriched prompts favored by systems is essential for the user experience. Such a prompt refinement process is analogous to translating the prompt from “user languages” into “system languages”. However, the scarcity of such parallel corpora makes it difficult to train a prompt refinement model. Inspired by zero-shot machine translation techniques, we introduce Prompt Refinement with Image Pivot (PRIP). PRIP innovatively uses the latent representation of a user-preferred image as an intermediary “pivot” between the user and system languages. It decomposes the refinement process into two data-rich tasks: inferring representations of user-preferred images from user languages and subsequently translating image representations into system languages. Thus, it can leverage abundant data for training. Extensive experiments show that PRIP substantially outperforms a wide range of baselines and effectively transfers to unseen systems in a zero-shot manner.

Legal case retrieval is a critical process for modern legal information systems. While recent studies have utilized pre-trained language models (PLMs) based on the general domain self-supervised pre-training paradigm to build models for legal case retrieval, there are limitations in using general domain PLMs as backbones. Specifically, these models may not fully capture the underlying legal features in legal case documents. To address this issue, we propose CaseEncoder, a legal document encoder that leverages fine-grained legal knowledge in both the data sampling and pre-training phases. In the data sampling phase, we enhance the quality of the training data by utilizing fine-grained law article information to guide the selection of positive and negative examples. In the pre-training phase, we design legal-specific pre-training tasks that align with the judging criteria of relevant legal cases. Based on these tasks, we introduce an innovative loss function called Biased Circle Loss to enhance the model’s ability to recognize case relevance in fine grains. Experimental results on multiple benchmarks demonstrate that CaseEncoder significantly outperforms both existing general pre-training models and legal-specific pre-training models in zero-shot legal case retrieval. The source code of CaseEncoder can be found at https://github.com/Anonymous-EMNLP2023/CaseEncoder.

Being able to induce word translations from non-parallel data is often a prerequisite for cross-lingual processing in resource-scarce languages and domains. Previous endeavors typically simplify this task by imposing the one-to-one translation assumption, which is too strong to hold for natural languages. We remove this constraint by introducing the Earth Mover’s Distance into the training of bilingual word embeddings. In this way, we take advantage of its capability to handle multiple alternative word translations in a natural form of regularization. Our approach shows significant and consistent improvements across four language pairs. We also demonstrate that our approach is particularly preferable in resource-scarce settings as it only requires a minimal seed lexicon.