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

A research team has introduced ACAL (Adaptive Collaboration of Argumentative LLMs), a novel neuro-symbolic framework designed to bring structured transparency and contestability to AI-powered legal reasoning. The system directly addresses the critical limitation of current LLM approaches like Chain-of-Thought (CoT) and Retrieval-Augmented Generation (RAG), which often produce unstructured, black-box explanations that are difficult to verify or challenge.

Technically, ACAL integrates adaptive multi-agent collaboration with an Arena-based Quantitative Bipolar Argumentation Framework (A-QBAF). It dynamically deploys teams of specialized AI agents to construct competing legal arguments. A core 'clash resolution' mechanism then adjudicates these conflicting claims, and an 'uncertainty-aware escalation' process handles borderline cases. The most significant innovation is its Human-in-the-Loop (HITL) contestability workflow, which allows users—like lawyers or judges—to directly audit, interrogate, and even modify the underlying reasoning graph that leads to a judgment, enabling them to influence the final outcome.

In empirical evaluations on the LegalBench benchmark, ACAL demonstrated superior performance, outperforming strong baselines built on both the Gemini-2.5-Flash-Lite and the more capable Gemini-2.5-Flash architectures. This indicates the framework's effectiveness isn't just about raw accuracy but about achieving a better balance between efficient prediction and the structured, verifiable reasoning required in legal domains. The implementation is publicly available, signaling a move toward more accountable and interactive AI systems for high-stakes decision-making.