AgentRewardBench (Screenshot) - WebArena: leaderboard
AgentRewardBench judge precision (%) on WebArena trajectories, with the authors' simplified judge reading the final screenshot: the share of trajectories it calls successful that experts also marked successful; higher is better (McGill NLP, 2025).
Metric: Precision (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-11-20) | 69.9 | #864 |
| 2 | Claude 3.7 Sonnet | 69.3 | #455 |
| 3 | GPT-4o Mini (2024-07-18) | 66.9 | #1512 |
| 4 | Qwen 2.5 VL 72B Instruct | 62.9 | #845 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=agentrewardbench-screenshot-webarena · How It Works · Data refreshed daily, snapshot 2026-11-12.