LePhantomCite - Non-Agentic: leaderboard
LePhantomCite (Who Checks the Citations?, 2026) tests detection of hallucinated legal citations in brief excerpts with injected errors such as non-existent cases, wrong pincites, misquotes and misrepresented holdings. This board is the F1 score in percent of flagged segments on 390 test excerpts in the non-agentic setting, with no case-law database lookups; higher is better.
Metric: F1 (%) of predicted hallucinated segments against the injected ground-truth segments (a prediction matches when either span contains the other), non-agentic setting: the model reads a legal brief excerpt with the agent prompt minus the belief-update and action-selection sections, with no database lookups; 390 test excerpts; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 45.2 | #167 |
| 2 | Gemini 2.5 Flash | 32 | #453 |
| 3 | Qwen 3 8B | 20.7 | #1830 |
| 4 | GPT-OSS-120B | 20 | #700 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=lephantomcite-non-agentic · How It Works · Data refreshed daily, snapshot 2026-11-13.