ZwZ2 Paper - HR-Bench 4K: leaderboard
HR-Bench 4K (Wang et al.), a fine-grained perception benchmark in the paper's grouping: perception questions on 4K-resolution images. ZwZ2 paper (arXiv 2610.37037v1, Table 1): the authors evaluated the base models Qwen3.5-4B, Qwen3.5-9B and Qwen3.6-27B and the ZwZ2-4B/9B/27B models they trained from them with GRPO in non-thinking mode. The table notes say only that all models were evaluated in low-thinking or non-thinking modes, without naming the mode per model; decoding, prompts, input resolution and evaluation framework are not reported. Accuracy (%), higher is better.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 9B | 84.88 | #391 |
| 2 | Qwen 3.6 27B | 84.25 | #182 |
| 3 | Qwen 3.5 4B | 82.37 | #514 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=zwz2-paper-hr-bench-4k · How It Works · Data refreshed daily, snapshot 2026-11-05.