Verified, then logged.
Sun 22 Nov 2026 · 00:33 KST / 15:33 UTC한국어 | ENSearch /

Breaking01 / AI

Legal AI training method balances reasoning and answers

The method balances learning from reasoning and final answers, then adjusts auxiliary rewards according to their effect on the final evaluation.

A smartphone on speakerphone rests on a living-room coffee table while a television and a family member seen from behind blur in the background. AI-generated image

Researchers proposed a two-stage training method for legal reasoning tasks in which reasoning is much longer than the final answer.

The paper, published on arXiv on 13 Nov, describes a first stage called JurisCoT-SFT. It adjusts for text length to balance learning from reasoning and final answers. This addresses the weakening of the answer's learning signal when reasoning takes up much more of the training text.

In the second stage, the researchers adjust auxiliary rewards according to their effect on the final evaluation. These rewards are extra scores used to guide training. The method evaluates them and removes them when needed.

The researchers reported gains on selected tasks from ChJGBench, LawBench and LegalBench, benchmark sets used to evaluate legal AI. Their reported results showed improvements in reasoning quality and final-answer accuracy, while their models remained competitive with larger models.

The paper distinguishes the method from earlier approaches by addressing both the weight given to answer learning and the management of rewards. It identifies diluted answer-learning signals as a problem in supervised fine-tuning, which trains models on examples, and costly reward design and stalled performance as problems in reinforcement learning, which trains models using reward scores.

The researchers did not disclose task-level scores or the names of comparison models in the abstract.