LegalEvolver
A query optimization framework for legal judgment prediction (LJP). Through heuristic seed queries, rubric-guided multi-round self-evolution, and reinforcement learning, it generates retrieval queries rich in legal information to improve the performance of legal AI agents.
Timeline (last 90 days)
LegalEvolver is a query optimization framework specifically designed to generate legally information-rich queries for LJP-oriented retrieval.
Directly sampling high-quality optimized queries and their corresponding search trajectories is very difficult; relying solely on unguided high-temperature random sampling is not only inefficient but also limited by the model's initial capabilities.
LegalEvolver first optimizes queries through heuristic methods, uses them as seeds, and then generates search trajectories based on these seeds.
LegalEvolver summarizes the experience of each rollout into scoring rubrics and uses these rubrics to guide a multi-round iterative self-evolution process; the model continuously optimizes the scoring rubrics and uses the optimized rubrics to guide reinforcement learning (RL).
Comparative experiments in the paper show that evolving scoring rubrics are more effective than fixed rubrics.
Compared with the approach of directly applying supervised fine-tuning (SFT) followed by reinforcement learning, the rubric-guided method performs better.
In the context of this paper, scoring rubrics are not a manually pre-fixed scoring table, but rather a set of natural-language criteria that the model summarizes from sampling experience and uses to judge the quality of queries.
The self-evolution mechanism enables the system to surpass the performance ceiling set by the initial rubrics or the initial model capabilities.
The paper leaves open questions: how to guarantee the quality of the scoring rubrics, and whether the self-evolution process introduces bias.
All Facts (9)
LegalEvolver is a query optimization framework specifically designed to generate legally information-rich queries for LJP-oriented retrieval.
50%UnverifiedDirectly sampling high-quality optimized queries and their corresponding search trajectories is very difficult; relying solely on unguided high-temperature random sampling is not only inefficient but also limited by the model's initial capabilities.
50%UnverifiedLegalEvolver first optimizes queries through heuristic methods, uses them as seeds, and then generates search trajectories based on these seeds.
50%UnverifiedLegalEvolver summarizes the experience of each rollout into scoring rubrics and uses these rubrics to guide a multi-round iterative self-evolution process; the model continuously optimizes the scoring rubrics and uses the optimized rubrics to guide reinforcement learning (RL).
50%UnverifiedComparative experiments in the paper show that evolving scoring rubrics are more effective than fixed rubrics.
50%UnverifiedCompared with the approach of directly applying supervised fine-tuning (SFT) followed by reinforcement learning, the rubric-guided method performs better.
50%UnverifiedIn the context of this paper, scoring rubrics are not a manually pre-fixed scoring table, but rather a set of natural-language criteria that the model summarizes from sampling experience and uses to judge the quality of queries.
50%UnverifiedThe self-evolution mechanism enables the system to surpass the performance ceiling set by the initial rubrics or the initial model capabilities.
50%UnverifiedThe paper leaves open questions: how to guarantee the quality of the scoring rubrics, and whether the self-evolution process introduces bias.
50%