HalluWorld - Grid Uncertainty: leaderboard

Uncertainty hallucination in HalluWorld-Grid (2026), MiniGrid worlds whose ground truth comes from the environment state: the rate of hallucinated answers to questions where the evidence is insufficient and the model should abstain. Percent; lower is better.

Metric: Hallucination rate (%) on the uncertainty probes (abstaining when the evidence is insufficient, 5 levels) of HalluWorld-Grid (MiniGrid worlds whose ground truth comes from the environment state), micro-averaged over levels and serializers, temperature 0 where supported, thinking models at medium effort with up to 16k thinking tokens, every model at most 256 answer tokens; lower is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GLM-5Non-reasoning0#269
2GPT-5.5Medium0.1#54
3O30.3#237
4GPT-4o0.5#751
5O4 Mini0.9#368
6Claude Sonnet 4.61.1#173
7Kimi K2.6Non-reasoning1.9#192
8Claude Opus 4.63.4#113
9DeepSeek V3 (0324)4.6#715
10GPT-4o Mini5.1#1667
11GPT-5.4 MiniNon-reasoning5.7#406
12O3 Mini5.9#546
13Qwen 3 30B A3BNon-reasoning13.2#1156

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=halluworld-grid-uncertainty · How It Works · Data refreshed daily, snapshot 2026-11-21.