CSR-Bench (Hallucination): leaderboard
Metric: Alignment accuracy (%) averaged over the four question forms (original, wrong premise, distractor text, modality conflict); 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 2028. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 51.3 | #77 |
| 2 | Qwen 3 VL 8B Instruct | 41.4 | #401 |
| 3 | Qwen 3 VL 8B (Thinking) | 40.2 | |
| 4 | InternVL3-14B | 39.8 | #494 |
| 5 | Qwen 3 VL 4B (Thinking) | 38.4 | #471 (Qwen 3 VL 4B) |
| 6 | Qwen 3 VL 4B Instruct | 37.6 | #506 |
| 7 | Gemini 2.5 Flash | 37.2 | #237 |
| 8 | InternVL3-8B | 35.7 | #606 |
| 9 | Qwen 2.5 VL 7B Instruct | 34.3 | #643 |
| 10 | Seed-1.6 | 32.9 | #257 |
| 11 | Gemma 3 12B (IT) | 30.1 | #655 |
| 12 | Llama 3.2 11B Instruct | 26.2 | #1112 |
| 13 | Gemma 3 4B (IT) | 23.9 | #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 · How It Works · Data refreshed daily, snapshot 2026-10-11.