In the legal domain, LLMs can improve reliability and interpretability by generating explicit reasoning traces. However, training LLMs to produce high-quality reasoning traces remains challenging. Existing supervised fine-tuning methods struggle when reasoning traces are significantly longer than final answers, as the learning signal for the answer becomes diluted. Meanwhile, reinforcement learning methods such as Group Relative Policy Optimization also have several limitations, including costly reward design, performance plateaus, and reward hacking. We therefore propose a two-stage training framework. In Stage I, JurisCoT-SFT employs a length-normalized dual objective to balance learning signals between reasoning traces and final answers. In Stage II, Lifecycle-Aware Backtrackable Policy Optimization dynamically evaluates and retires auxiliary reward signals based on their impact on final evaluation objectives, enabling efficient reward utilization. Our framework substantially improves legal reasoning performance on relevant tasks selected from multiple legal benchmarks, including ChJGBench, LawBench, and LegalBench. Experimental results demonstrate consistent improvements in both reasoning trace quality and final answer accuracy, while our models remain competitive with larger models.

Large language models have demonstrated strong performance on general-purpose tasks but often fail to satisfy the accuracy requirements of knowledge-intensive domains such as law, medicine, and finance. Complex domain-specific generation is inherently compositional, involving multiple atomic skills such as reasoning, knowledge grounding, and numerical computation that are frequently interleaved at the token level. Existing domain adaptation methods typically train these heterogeneous skills jointly within a single objective, which makes it difficult for models to reliably coordinate multiple skills when solving complex tasks. In this work, we explicitly incorporate atomic skills into domain-specific model training and propose SplitThenMerge, a framework that decomposes domain competence into atomic skills, trains them independently, and composes them dynamically during generation. SplitThenMerge adopts a token-level sparse Mixture-of-Experts architecture to enable fine-grained skill routing and coordination while implementing each skill as a lightweight LoRA expert to achieve parameter-efficient specialization. Experimental results demonstrate that our method consistently achieves superior performance in both legal and medical domains under the same training parameter budget.

Artificial intelligence has become increasingly prevalent in the legal domain. However, LegalAI systems often struggle with vague user queries that lack essential legal details, leading to suboptimal performance in practical applications. To address this challenge, we propose FactFiller, a novel approach that dynamically generates questionnaires to help users refine their input queries. Our method leverages an iterative training process that collects valuable questionnaires, eliminating the need for human annotation. Additionally, we introduce a "case-law-quiz” cascading retrieval process, ensuring that the generated questions and answer options are directly linked to specific legal provisions. Through the user study and the downstream task experiments, we demonstrate that FactFiller, while remaining easy for non-experts to understand, not only improves the completeness of queries but also ensures the performance of various domain-specific models in downstream legal tasks.

With the rapid development of LLMs, increasing attention has been paid to whether they can reliably generate complete and legally valid judicial documents. This calls for a benchmark capable of evaluating whether models can accurately apply statutes and resolve outcome-critical issues, including claims, charges, penalties, and monetary obligations. However, existing benchmarks typically focus on isolated legal tasks or coarse-grained evaluation. We introduce ChJGBench, a benchmark for generating the reasoning and dispositive sections of Chinese judgments across civil, criminal, and administrative cases. Built from 3,462 real-world judgments, ChJGBench provides a case-type-aware diagnostic evaluation framework comprising 13 metric types and adopts a hybrid paradigm. Experiments reveal no uniform advantage for legal-domain models over general-purpose models, with model strengths varying across diagnostic dimensions. Performance is generally stronger on criminal cases, while civil and administrative cases continue to present substantial challenges.

As Large Language Models (LLMs) are widely applied in various domains, the safety of LLMs is increasingly attracting attention to avoid their powerful capabilities being misused. Existing jailbreak methods create a forced instruction-following scenario, or search adversarial prompts with prefix or suffix tokens to achieve a specific representation manually or automatically. However, they suffer from low efficiency and explicit jailbreak patterns, far from the real deployment of mass attacks to LLMs. In this paper, we point out that simply rewriting the original instruction can achieve a jailbreak, and we find that this rewriting approach is learnable and transferable. We propose the Rewrite to Jailbreak (R2J) approach, a transferable black-box jailbreak method to attack LLMs by iteratively exploring the weakness of the LLMs and automatically improving the attacking strategy. The jailbreak is more efficient and hard to identify since no additional features are introduced. Extensive experiments and analysis demonstrate the effectiveness of R2J, and we find that the jailbreak is also transferable to multiple datasets and various types of models with only a few queries. We hope our work motivates further investigation of LLM safety. The code can be found at https://github.com/ythuang02/R2J/.