AgentRewardBench (Screenshot): leaderboard
AgentRewardBench (McGill NLP, 2025) tests whether an LLM judge can tell if a web agent succeeded, using 1,302 expert-annotated trajectories from AssistantBench, VisualWebArena, WebArena, WorkArena and WorkArena++. This board scores the authors' simplified judge given the final screenshot of each trajectory, as precision (%): the share of trajectories the judge calls successful that the experts also marked successful; higher is better.
Metric: Precision (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.7 Sonnet | 69.4 | #455 |
| 2 | GPT-4o (2024-11-20) | 68.1 | #864 |
| 3 | GPT-4o Mini (2024-07-18) | 64.5 | #1512 |
| 4 | Qwen 2.5 VL 72B Instruct | 64.5 | #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 · How It Works · Data refreshed daily, snapshot 2026-11-12.