LongJud-Bench - Position Consistency (First Instance): leaderboard
LongJud-Bench (LegalWorld, Zuo et al., 2026) makes a model the target lawyer in a simulated Chinese civil case, from client consultation to second-instance trial, with every other role played by Qwen3.5-Plus on a 100-case split. Position consistency is scored on the agent's first-instance trial statements, for consistency with its pleaded claims or defenses, by a Claude Sonnet 4.6 judge on a 0-10 rubric per trial phase (a phase with no required statement scores zero), normalized to percent; higher is better.
Metric: Position Consistency Score (%; Claude Sonnet 4.6 judge, 0-10 per trial phase on consistency of the first-instance trial statements with the pleaded claims or defenses, normalized). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K2.5 | 63 | #252 |
| 2 | GPT-5.2 | 62 | #172 |
| 3 | Qwen 3.5 Plus | 62 | #227 |
| 4 | GLM-4.7 | 61 | #365 |
| 5 | DeepSeek V4 Flash | 60 | #254 |
| 6 | Qwen 3.5 Flash | 53 | #443 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=longjud-bench-position-consistency-first-instance · How It Works · Data refreshed daily, snapshot 2026-11-12.