Large Language Models (LLMs) have become increasingly useful in legal practice, yet when non‑expert users provide underspecified consultations, existing models often generate irresponsible and low‑utility responses because the case details are insufficient for a definitive conclusion. To better assist laypeople, legal LLMs must proactively identify the dispute’s focal points and solicit crucial context from users. Current datasets mainly evaluate interaction quality and do not reveal how proactive interaction influences downstream claim‑level judgment prediction. To bridge this gap, we introduce LexPilot, a benchmark that links proactive legal interaction with claim‑level verdict prediction using authentic Chinese civil judgments. LexPilot consists of 432 real judgments covering 36 common causes of action, offering expert‑reviewed pairs of fully specified and underspecified consultation histories together with the corresponding court decisions. Experiments show that, under underspecified consultations, present LLMs struggle with proactive interaction, causing downstream judgment prediction to drop by about 23.84% on average. Further analysis pinpoints critical bottlenecks, highlighting the urgent need for user‑centered legal AI.
Review