ARM-Thinker Study - VL-RewardBench Hallucination: leaderboard
Agentic evaluation of supplied candidate answers, not answer generation. ARMBench-VL covers 1,499 fine-grained perception, instruction-following and document cases; Table 2 also reports VL-RewardBench and RewardBench-2. Preserve this judging role and the trained ARM-Thinker checkpoint; do not borrow the separately copied Table 3 peer scores.
Metric: Judgment Accuracy (%). Source: arxiv.org. Saturation forecast: Around January 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 8B Instruct | 71.3 | #414 |
| 2 | GPT-4o | 67.6 | #334 |
| 3 | InternVL3-8B | 51.3 | #611 |
| 4 | Qwen 2.5 VL 7B Instruct | 48.7 | #644 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #7 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=arm-thinker-study-vl-rewardbench-hallucination · How It Works · Data refreshed daily, snapshot 2026-10-16.