A new arXiv preprint introduces reward models designed to train language models to cite retrieved evidence when answering legal questions and to decline when that evidence is inadequate. The authors note that most existing reward models are not built for these high-stakes requirements. The work targets legal applications where unsupported or fabricated answers carry significant risk.
Researchers have introduced LexAgentHallu, a new benchmark for measuring how tool-augmented legal AI agents hallucinate. It uses a hierarchical structure to trace how errors in tool calls and reasoning cascade into fabricated case holdings and miscited legal authority. The benchmark aims to fill a gap left by existing legal evaluations that do not capture agentic workflows.
A new paper examines whether and how a defendant's courtroom statement shifts the verdicts of large language models simulating jurors in common-law trials. The authors focus on persuasion effects and potential bias in these simulated decisions, a setting where LLM-based modeling of human judgment had not previously been studied.
Researchers have proposed GANDR, an approach that audits legal question-answering systems at the level of individual claims instead of scoring whole answers. Each statement generated by the model is checked against the specific source it cites, so readers can verify a grounded answer piece by piece. The work targets a shortcoming of existing grounded-generation pipelines, which typically evaluate answers only in aggregate.