LetTokenBurnLexPilot: Benchmarking the Proactive Interaction Ability of Legal LLMs with Laypeople in Chinese Civil Case Consultation
LexPilot evaluates legal LLMs' proactive interaction and judgment prediction using 432 expert-reviewed pairs of complete and underspecified Chinese civil consultation histories.
$ npm i -g lettokenburn-cli$ ltb bench info lexpilot
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Claude Opus 5 | 0.7 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| GPT-5.6 Sol | 0.7 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Gemini 3.8 Flash | 0.7 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| LegalOne-8B | 0.7 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| GLM-5.3 | 0.6 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Kimi K3 | 0.6 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| MiniMax-M3 | 0.6 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| DeepSeek V4 Pro | 0.6 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| Qwen3.8 Max | 0.6 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
| LawLLM-7B | 0.4 | N/A | Independent | medium | Oct 16, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Constructed from 432 real-world Chinese civil judgments across 36 causes of action, LexPilot features expert-reviewed pairs of fully specified and underspecified consultation histories linked to court verdicts. It assesses Legal Completeness Assessment, Detail Completion Guidance, and Legal Judgment Prediction.
Scoring relies on downstream claim-level judgment prediction accuracy. Experiments show current LLMs struggle with proactive interactions, exhibiting a 23.84% average performance degradation compared to fully specified setups, highlighting critical bottlenecks in omission detection and guidance.