LegalEvolver: Score-Guided Self-Evolution Breaks Through Legal Judgment Prediction Bottlenecks

Legal judgment prediction (LJP) requires AI agents to perform multi-step reasoning and retrieve external legal evidence, but the performance of existing legal search agents is often limited by the quality of retrieval queries. They tend to construct queries directly from the user's initial request, ignoring potential legal issues and legally salient details. To address this problem, researchers have proposed LegalEvolver—a query optimization framework specifically designed to generate legally information-rich queries for LJP-oriented retrieval.

Query Quality: The Hidden Bottleneck of Legal AI Agents

As LLM agents gradually penetrate the legal domain, legal judgment prediction (LJP) is regarded as a high-value but highly difficult task. It requires the model not only to understand case facts but also to perform multi-step reasoning and retrieve external evidence such as legal statutes and precedents. However, existing legal search agents share a common problem: the quality of retrieval queries often directly determines the success or failure of the final prediction.

According to the analysis in the original paper, current methods typically construct queries directly from the user's initial request, ignoring potential legal issues and legally salient details. For example, a user may describe only the basic facts of a case, but legal reasoning needs to focus on key elements such as "subjective intent," "causation," and "timeliness." If the query does not cover these aspects, the retrieved evidence will be incomplete, which in turn affects the accuracy of judgment prediction.

The essence of this bottleneck is that query generation is treated as an isolated step rather than a process that requires optimization and iteration. LegalEvolver was proposed precisely to break this limitation.

Legal judgment prediction (LJP) is a classic task at the intersection of natural language processing and law, typically modeled as predicting court judgment outcomes (such as charges, statutes, and sentences) based on case fact descriptions. Unlike general question answering, LJP is highly specialized and structured: case facts often mix everyday narratives with legal elements, and the model needs to identify legal constitutive elements such as "subject, act, subjective state, causation, and harm" and match them to the corresponding legal norms. Traditional methods mostly use rule-based or statistical classification models, while in recent years agent-based methods built on large language models have begun to introduce external retrieval to compensate for the lack of legal knowledge in model parameters. Understanding the task definition of LJP helps to grasp why the quality of retrieval queries becomes the bottleneck of the entire pipeline.

Three Core Mechanisms of LegalEvolver

To solve the query quality problem, LegalEvolver designs a framework that combines heuristic optimization with reinforcement learning. Its core mechanisms can be summarized in the following three points:

1. Heuristic Query Optimization as Seeds

Directly sampling high-quality optimized queries and their corresponding search trajectories is very difficult. The paper points out that relying solely on unguided high-temperature rollouts is not only inefficient but also limited by the model's initial capabilities. LegalEvolver's approach is to first optimize queries through heuristic methods and use them as seeds, then generate search trajectories based on these seeds. This significantly improves the efficiency of trajectory generation and avoids blind exploration in a huge search space.

Rollout in reinforcement learning refers to the process in which a model executes a series of actions under a given state and collects rewards. In LegalEvolver, a rollout corresponds to the complete interaction sequence of "generating a query—executing retrieval—obtaining evidence—forming a prediction." High-quality rollouts are key material for reinforcement learning training, but random sampling directly in a huge action space often yields a large number of low-quality trajectories. Heuristic seed queries serve as a prior starting point for search: first use domain rules or templates to rewrite the initial request into a query closer to legal elements, then use this as the starting point for sampling. This approach is often called "warm start" or "guided exploration" in reinforcement learning and can significantly reduce sampling variance and improve sample efficiency.

2. Multi-Round Self-Evolution Guided by Rubrics

LegalEvolver summarizes the experience of each rollout into rubrics and uses these rules to guide a multi-round iterative self-evolution process. Specifically, the model continuously optimizes the rubrics and uses the optimized rules to guide reinforcement learning (RL). This iterative mechanism enables the model to surpass its initial performance ceiling.

Rubrics in the context of this paper are not manually predefined scoring tables, but rather a set of natural language criteria summarized by the model from sampling experience for judging the quality of queries, such as "the query should clearly point to the disputed statute" and "it should include time and jurisdictional elements." Using these rules as a source of reward signals for reinforcement learning belongs to an "AI feedback"-style reward modeling approach. Unlike RLHF, where a reward model is trained from human preferences, here the rules are iteratively generated and revised by the model itself. Multi-round self-evolution refers to evaluating and filtering trajectories with the current rules in each round, updating the policy, and then revising the rules based on new experience, forming a closed loop in which rules and policy co-evolve. This mechanism enables the system to surpass the performance ceiling set by the initial rules or initial model capabilities.

3. End-to-End Performance Validation

The paper conducts end-to-end evaluation on a wide range of benchmarks, and the results demonstrate the effectiveness of query optimization for LJP. Compared with directly applying supervised fine-tuning (SFT) followed by reinforcement learning, the rubric-guided method performs better; and the evolved rubrics further outperform fixed rubrics, confirming the gains brought by self-evolution.

Self-Evolution and Rubrics: Why Do They Work?

From a methodological perspective, the innovation of LegalEvolver lies in transforming "query optimization" from a one-time action into a continuous self-improvement loop. There are two key insights behind this:

• Queries in the legal domain require injection of domain knowledge: Legal issues are often implicit in factual descriptions, and the model needs to actively uncover potential legal disputes. Rubrics are essentially an explicit encoding of the knowledge of "what makes a good query," and through iterative optimization, the model gradually internalizes this knowledge.
• Self-evolution breaks through capability ceilings: The traditional SFT+RL pipeline relies on fixed human-annotated data or reward functions, and model performance is limited by the quality of the initial data. LegalEvolver, by letting the model summarize rubrics itself and continuously optimize them, forms a "self-improvement" loop. The comparative experiments in the paper confirm that evolving rubrics are more effective than fixed rules, indicating that the model indeed acquires new capabilities during iteration.

Implications for Legal AI Applications

Although LegalEvolver is a research framework, its approach has direct reference value for deploying legal AI products. Many current legal retrieval tools or judgment prediction systems still remain at the stage of "user input equals query," ignoring the room for query reconstruction and optimization. LegalEvolver suggests that:

• Query optimization can be embedded as an independent, learnable module in legal AI pipelines;
• Rubrics are a lightweight way to inject knowledge, more sustainable than large-scale human annotation;
• Self-evolution mechanisms help models adapt to dynamic changes in the legal domain (such as new regulations and new precedents) without frequent retraining.

Of course, the paper also leaves open questions: How can the quality of rubrics be guaranteed? Will the self-evolution process introduce bias? These require further exploration in future research.

Conclusion

Through heuristic seed queries, iterative self-evolution guided by rubrics, and end-to-end experimental validation, LegalEvolver provides a new path for query optimization in legal judgment prediction. Its core contribution lies in revealing the key impact of query quality on LJP performance and providing a scalable self-improvement solution. For researchers and developers concerned with the deployment of AI agents in the legal domain, this work is worth reading in depth.

One-line Summary

LegalEvolver optimizes retrieval queries for legal judgment prediction through heuristic seed queries and rubric-guided self-evolution.

Paragraph Summary

The article introduces LegalEvolver—a query optimization framework for legal judgment prediction (LJP). Existing legal search agents typically construct retrieval queries directly from the user's initial request, ignoring potential legal issues and legally salient details, which leads to incomplete retrieved evidence and limited prediction performance. LegalEvolver proposes three core mechanisms: first, optimize queries through heuristic methods as seeds to improve the sampling efficiency of search trajectories; second, summarize the experience of each rollout into rubrics and use them to guide multi-round iterative reinforcement learning self-evolution, breaking through the model's initial capability ceiling; finally, conduct end-to-end validation on a wide range of benchmarks. Experiments show that the rubric-guided method outperforms direct SFT+RL, and evolved rubrics outperform fixed rules. The article also discusses the implications of this method for deploying legal AI products and open questions.