Xiaoyang Yuan
Publications
MAP-Law: Coverage-Driven Retrieval Control for Multi-Turn Legal Consultation
Legal consultation is a high-stakes, knowledge-intensive task that requires agents to identify relevant legal issues, retrieve authoritative support, and determine when evidence is sufficient for a recommendation. Although retrieval-augmented generation has improved grounding in legal question answering, many multi-turn legal agents still rely on fixed retrieval depth or coarse heuristic control. This often leads to either insufficient support for key legal elements or excessive retrieval that increases context burden and weakens answer focus. We propose MAP-Law, a coverage-driven framework for retrieval control in multi-turn legal consultation. MAP-Law models consultation as a controlled retrieval process over a joint structured state consisting of issue nodes, legal element nodes, and evidence nodes. After each retrieval round, the agent computes Element Coverage, Evidence Coverage, and Marginal Gain, and uses these signals to decide whether to continue retrieval, redirect the search, or generate the final response. In this way, MAP-Law turns stopping from a fixed hyperparameter into an interpretable and auditable decision aligned with legal argumentative structure. Experiments on a self-constructed dataset of 50 cases across eight labor-law scenarios show that MAP-Law with DeepSeek as the action selector achieves an Element Coverage of 0.860 using only 2.9 retrieval rounds and 5.8 evidence pieces on average. Compared with a fixed seven-round baseline, it reduces evidence volume by over 80% and retrieval rounds by 58%. Ablation results further confirm the independent contributions of coverage-driven stopping, joint graph representation, and LLM-based action selection.
RISER: Orchestrating Latent Reasoning Skills for Adaptive Activation Steering
Recent work on domain-specific reasoning with large language models (LLMs) often relies on training-intensive approaches that require parameter updates. While activation steering has emerged as a parameter efficient alternative, existing methods apply static, manual interventions that fail to adapt to the dynamic nature of complex reasoning. To address this limitation, we propose RISER (Router-based Intervention for Steerable Enhancement of Reasoning), a plug-and-play intervention framework that adaptively steers LLM reasoning in activation space. RISER constructs a library of reusable reasoning vectors and employs a lightweight Router to dynamically compose them for each input. The Router is optimized via reinforcement learning under task-level rewards, activating latent cognitive primitives in an emergent and compositional manner. Across seven diverse benchmarks, RISER yields 3.4-6.5% average zero-shot accuracy improvements over the base model while surpassing CoT-style reasoning with 2-3x higher token efficiency and robust accuracy gains. Further analysis shows that RISER autonomously combines multiple vectors into interpretable, precise control strategies, pointing toward more controllable and efficient LLM reasoning.