Adaptive Collaboration of Arena-Based Argumentative LLMs for Explainable and Contestable Legal Reasoning

This repository contains the official implementation and evaluation code for the paper:

Adaptive Collaboration of Arena-Based Argumentative LLMs for Explainable and Contestable Legal Reasoning [Paper]

---

Overview

Installation

Create environment

We recommend using conda to set up the environment.

conda create -n env python=3.11
conda activate env
pip install -r requirements.txt

API Configuration

Create a .env file in the root directory and configure the required API keys.

# Gemini API
GOOGLE_API_KEY="YOUR_GOOGLE_API_KEY"

# Azure OpenAI (for RAG setup)
AZURE_API_KEY=""
AZURE_ENDPOINT=""

⚠️ Important:

Do NOT commit your .env file or API keys to the repository.

Running the Framework

To run the main experimental pipeline:

python main.py

Customization and Extensions

Ablation Studies

• You can modify the base_adjustment parameter in qbaf_scorer.py to perform ablation studies on the argument strength adjustment mechanism.

Prompt Engineering

• Agent prompts are defined in node.py.

• The current implementation supports two legal reasoning tasks used in the paper.

• You can extend the framework to other legal tasks by defining new prompts and agent roles.