DLawBench: Evaluating LLMs Through Multi-Turn Legal Consultation

2026-06-11

AI Summary

This study addresses a critical gap in current legal AI evaluation frameworks, which overlook the capacity of large language models to proactively elicit key facts and adapt to diverse client behaviors during multi-turn attorney–client consultations. To bridge this gap, the work introduces a novel typology of client behaviors—cooperative, dependent, withdrawn, and adversarial—and constructs a structured, case-based benchmark for multi-turn legal dialogues grounded in real-world scenarios. It proposes a consultation-centric paradigm for evaluating legal reasoning, integrating factual completeness, questioning quality, and dispute resolution efficacy as core metrics. A systematic assessment of 26 leading models reveals that even the top-performing model (GPT-5.5) achieves only a modest score of 0.562, highlighting significant performance degradation in high-guidance contexts and a tendency toward excessive client appeasement.

Abstract

Lawyer-client consultation is a critical starting point for legal services. Effective legal assistance hinges on eliciting sufficient and truthful information from clients in order to devise strategies that best protect their interests. This task requires Large Language Models (LLMs) not only to perform robust legal reasoning, but also to strategically elicit material facts through multi-turn interactions and effectively guide clients with diverse personalities. Yet existing legal benchmarks overlook this interactive capability. To fill this gap, we introduce DLawBench, a diagnostic benchmark for real-world legal consultation. Drawing on realistic client behavior, we characterize lawyer-client interactions into four types: Cooperative, Dependent, Withdrawn, and Adversarial. Using dialogues grounded in real cases, DLawBench evaluates whether LLMs can effectively conduct legal consultation under realistic conditions. DLawBench comprises 461 cases from Chinese and U.S. law, 5,532 paired fact entries, 3,411 inquiry rubrics, and 3,348 issue-resolution rubrics, and evaluates 26 representative LLMs. Systematic experiments show substantial headroom: the best-performing model, GPT-5.5, achieves only 0.562 on consultation-grounded legal reasoning. More importantly, DLawBench exposes both sycophancy in legal consultation and a paradox: models perform worse when clients need guidance most.

Problem

• legal consultation
• multi-turn interaction
• large language models
• client behavior
• legal reasoning

Innovation

• multi-turn legal consultation
• interactive legal benchmark
• client behavior typology
• consultation-grounded reasoning
• LLM evaluation