ReactBench (Hallucination) - Relational Erasure (CoT): leaderboard

Accuracy in percent on ReactBench (2026) questions about images where an object or relation was edited out, with a step-by-step reasoning prompt; an incorrect answer counts as a hallucination. Higher is better.

Metric: Accuracy (%) on Relational Erasure (19,140 questions on 1,595 images where an object or relation was edited out) with a step-by-step reasoning prompt, hallucination-inducing questions (multiple choice, short answer, yes/no) about images edited by Qwen-Image-Edit-2511; an incorrect answer counts as a hallucination; temperature 0.2, at most 1,024 tokens, LLM answer matching; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 VL 32B (Thinking)60.7#573
2Qwen 3 VL 32B Instruct58.9#606
3ReactBench MiMo-VL-7B (checkpoint unspecified)55.2
4Qwen 3 VL 4B50.3#1151
5Qwen 2.5 VL 7B Instruct49.3#1902
6InternVL3-8B48.2#1782
7Qwen 3.5 9B47.4#828
8Qwen 2.5 VL 72B Instruct47.2#841
9Qwen 3.5 27B46.4#375
10InternVL2.5-26B44.8#2156
11ReactBench LLaVA-v1.6-7B (checkpoint unspecified)43.8
12ReactBench LLaVA-OV-7B (checkpoint unspecified)43.4
13InternVL2.5-8B33#2370

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

Interactive version: theaggregate.ai/benchmark?slug=reactbench-hallucination-relational-erasure-cot · How It Works · Data refreshed daily, snapshot 2026-11-21.