CSR-Bench (Hallucination) - Wrong Premise: leaderboard
Metric: Alignment accuracy (%) when a false description that contradicts the image is prepended to the question; CSR-Bench's hallucination subset: 1,411 images, each asked an original image-grounded question and three misleading variants, scored by a GPT-5-nano judge (1 when the response gives the image-grounded target or rejects the miscategorized request), temperature 0, 8,192-token cap; higher is better. Source: arxiv.org. Saturation forecast: Around July 2028. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 58.7 | #77 |
| 2 | Qwen 3 VL 8B (Thinking) | 51.3 | |
| 3 | InternVL3-14B | 49.8 | #494 |
| 4 | Qwen 3 VL 4B (Thinking) | 48.3 | #471 (Qwen 3 VL 4B) |
| 5 | Qwen 3 VL 8B Instruct | 48.1 | #401 |
| 6 | Gemini 2.5 Flash | 46.2 | #237 |
| 7 | InternVL3-8B | 45.3 | #606 |
| 8 | Qwen 3 VL 4B Instruct | 44.3 | #506 |
| 9 | Seed-1.6 | 42.1 | #257 |
| 10 | Qwen 2.5 VL 7B Instruct | 40 | #643 |
| 11 | Gemma 3 12B (IT) | 36.2 | #655 |
| 12 | Llama 3.2 11B Instruct | 35.2 | #1112 |
| 13 | Gemma 3 4B (IT) | 33.4 | #971 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=csr-bench-hallucination-wrong-premise · How It Works · Data refreshed daily, snapshot 2026-10-11.