Delulu - Method Hallucination Detection: leaderboard

LLMs as judges on the 461 Delulu (2026) fill-in-the-middle samples whose hallucinated completion contains calls to methods that do not exist: the judge must accept the real completion and reject the hallucinated one. The score is the both-correct rate (%); higher is better.

Metric: Both-correct rate (%) on the 461 Delulu samples whose hallucination is calls to methods that do not exist, as an LLM judge: the judge sees each sample twice, once with the gold completion and once with the hallucinated one, and must accept the first and reject the second (both-correct rate, %), with chain-of-thought and a binary verdict; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Opus 4.594.6#165
2GPT-5.2 Codex92.7#186
3GPT-5.484.2#138
4Claude Sonnet 4.678.4#179
5GPT-5.4 Mini78.2#409
6Claude Haiku 4.577.1#587
7GPT-4.1 Mini73#799
8GPT-4o Mini54.6#1680

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=delulu-method-hallucination-detection · How It Works · Data refreshed daily, snapshot 2026-11-22.