Test-Time Agent Evolution for Long-Horizon Legal Reasoning
A training-free legal agent adapts at test time using memory evolution and rubric-aligned collaboration.
The paper studies training-free test-time adaptation for long-horizon legal reasoning, where case facts, evidence, and procedures vary and decisions span multiple roles. Test-Time Memory Evolution retrieves experience from earlier cases, adapts it to the current context, and consolidates it for later decisions. Rubric-Aligned Collaboration verifies and revises role-specific actions against behavioral and procedural requirements. Experiments on J1-EVAL and LegalWorld across five backbone models show consistent gains over reasoning and agent baselines at moderate interaction and compute cost.
- Training-free agents adapt from prior cases without parameter updates.
- Test-Time Memory Evolution retrieves, adapts, and consolidates experience.
- Rubric-Aligned Collaboration checks actions against procedural requirements.
- Gains reported on J1-EVAL and LegalWorld across five backbone models.
- Ablations show complementary benefits for adaptation and cross-role coordination.
From the original article
1 min read at arxiv.orgLegal intelligence aims to support reliable decision-making across long-horizon legal processes involving evolving case states and multiple roles. However, real-world legal deployment exhibits substantial case heterogeneity in facts, evidence, and procedural contexts, exposing the limitations of static agent strategies.
By Haotian Chen · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.