Enhancing Judgment Document Generation via Agentic Legal Information Collection and Rubric-Guided Optimization

Imagine a judge's job as writing a very strict, high-stakes recipe for a legal decision. To write this "recipe" (the judgment document), the judge needs two things:

1. The Right Ingredients: They must find the exact laws and past court cases (statutes and precedents) that apply to the specific crime.
2. The Perfect Cooking Method: They must mix those ingredients with the facts of the case to create a logical, professional, and legally sound final dish.

Currently, computers trying to do this often make two big mistakes: they grab the wrong ingredients (hallucinating laws that don't exist) or they cook them together in a way that makes no logical sense.

The paper introduces Judge-R1, a new computer system designed to fix these problems. Think of it as a "Super-Chef" that uses a two-step training process to become the best legal writer possible.

Step 1: The "Smart Shopper" (Agentic Legal Information Collection)

Before writing, the system needs to find the right laws.

• The Old Way: Imagine a shopper who just types a vague description like "bad thing happened" into a search engine. They get back a huge pile of results, many of which are irrelevant or slightly wrong.
• The Judge-R1 Way: This system uses a Smart Shopper Agent. Instead of just searching, this agent acts like a detective. It breaks the messy story of the crime down into specific questions (e.g., "What is the law about theft?" "What is the penalty for this specific amount?"). It then goes to multiple "stores" (legal databases) to find the exact statutes and past cases needed.
• The Result: It filters out the noise and brings back a perfect, high-quality basket of legal evidence, ensuring the judge has the right "ingredients" before they start cooking.

Step 2: The "Strict Food Critic" (Rubric-Guided Optimization)

Once the system has the ingredients, it needs to write the judgment.

• The Old Way (Supervised Fine-Tuning): Imagine teaching a student to write by showing them 1,000 examples of good essays. The student learns to copy the style and format perfectly. However, they might still make logical errors or invent facts because they are just mimicking the look of the essay, not understanding the rules.
• The Judge-R1 Way: After the initial style training, the system enters a "Reinforcement Learning" phase. Here, it plays a game where it writes many different versions of the judgment.
  • A Strict Food Critic (a digital rubric) tastes every version.
  • The critic doesn't just look at how fancy the words are. It checks: Did you use the right law? Is the math on the fine correct? Is the logic sound? Did you repeat yourself too much?
  • If the judgment is legally wrong, the system gets a "bad score." If it's perfect, it gets a "good score."
  • The system learns from these scores, constantly adjusting its "recipe" until it consistently produces judgments that are not just well-written, but legally accurate and logically airtight.

The Final Result

The paper tested this "Super-Chef" against other top computer systems using a standard legal test called JuDGE.

• Finding the Ingredients: Judge-R1 found the correct laws much better than previous systems, which often missed key details or grabbed irrelevant ones.
• Writing the Judgment: The final documents produced by Judge-R1 were more accurate in their legal conclusions, had fewer made-up citations, and followed the strict logical structure required by courts better than the competition.

In short, Judge-R1 doesn't just try to sound like a lawyer; it uses a smart search strategy to find the truth and a strict scoring system to ensure the logic holds up, resulting in a much more reliable tool for generating legal judgments.