Legal agent errors mapped in new multi-step hallucination benchmark
LexAgentHallu: A Hierarchical Benchmark for Profiling Hallucinations in Legal Agents
Artificial Intelligence
Summary
Large language models used as legal assistants sometimes make mistakes where wrong facts or fake citations appear in their answers. Existing tests only check if the final answer is right or wrong, missing these errors along the way. The authors created LexAgentHallu, a new tool that tracks when and how these errors happen during multi-step reasoning in legal tasks. They tested many legal AI systems and found patterns in these mistakes that older tests could not see. This tool can help improve legal AIs by showing exactly where they go wrong.
What this means in practice
- •For legal technology developers: Diagnose and fix error patterns in AI legal assistants by using detailed multi-step hallucination data.
- •For software quality assurance teams: Test complex AI workflows with fine-grained metrics to identify exactly where reasoning and tool-call mistakes happen.
Authors
Yujin Zhou, Mingxuan Zheng, Chuxue Cao, Huang Yidan, Jiale Chen, Yike Guo, Sirui Han
Abstract
As large language models are increasingly deployed as tool-augmented legal agents, they introduce agentic hallucinations where tool-call and reasoning errors cascade into fabricated holdings and miscited authority. However, existing legal benchmarks evaluate only single-turn QA with outcome-level metrics, while agentic hallucination benchmarks lack legal-specific diagnostic capability. Neither answers to what extent and how a legal agent hallucinates along its trajectory. To address these limitations, we introduce LexAgentHallu, a legal agentic hallucination benchmark designed to evaluate to what extent and how legal agents fail along multi-step trajectories. Built through a four-stage expert-in-the-loop pipeline, LexAgentHallu contains 3414 instances across 17 legal categories and 6 task types. Each instance is annotated under a dual-layer hallucination taxonomy of 7 high-level categories and 27 fine-grained subclasses, covering both substantive errors and agent-procedural failures. We further design fine-grained metrics that quantify to what extent and localize how each failure occurs along an agent's execution path. Our evaluation across 18 proprietary and open-source agents uncovers a Right-Answer-Wrong-Reason effect and reveals that hallucination subclasses cluster rather than scatter, forming distinct agentic framework, legal task, and category profiles. These findings, invisible to outcome-level evaluation, validate the diagnostic power of LexAgentHallu for evaluating agentic hallucination in law.