AgentProcessBench - HotPotQA Step Accuracy: leaderboard
Metric: StepAcc (%): share of assistant steps, pooled over trajectories, whose correct (+1), neutral (0) or incorrect (-1) label from the judge matches the expert annotation, on the 250 HotPotQA trajectories (multi-hop question answering with a Wikipedia retriever) of AgentProcessBench's 1,000 human-annotated tool-use trajectories (250 per source benchmark, rollouts of five agent models); higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) (Non-reasoning) | 76.7 | #78 (Gemini 3 Flash (Preview)) |
| 2 | DeepSeek V3.2 (Non-reasoning) | 74.4 | #198 (DeepSeek V3.2) |
| 3 | GPT-5.2 (Medium) | 72.3 | #105 (GPT-5.2) |
| 4 | GPT-5.2 (Non-reasoning) | 72.1 | #105 (GPT-5.2) |
| 5 | Kimi K2.5 (Non-reasoning) | 71.7 | #139 (Kimi K2.5) |
| 6 | GPT-5.2 Instant | 71.4 | #205 |
| 7 | Kimi K2.5 (Thinking) | 70.6 | #139 (Kimi K2.5) |
| 8 | DeepSeek V3.2 (Thinking) | 70.4 | #198 (DeepSeek V3.2) |
| 9 | Qwen 3 30B A3B 2507 (Thinking) | 70 | #366 (Qwen 3 30B A3B 2507) |
| 10 | Qwen 3 4B 2507 (Thinking) | 66.9 | #525 (Qwen 3 4B 2507) |
| 11 | Qwen 3 4B 2507 Instruct | 66.6 | #745 |
| 12 | Qwen 3 30B A3B 2507 Instruct | 65.8 | #464 |
| 13 | Qwen 3 8B (Non-reasoning) | 60.4 | #667 (Qwen 3 8B) |
| 14 | Qwen 3 4B (Reasoning) | 60.1 | #823 (Qwen 3 4B) |
| 15 | Qwen 3 8B (Thinking) | 59.7 | #667 (Qwen 3 8B) |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=agentprocessbench-hotpotqa-step-accuracy · How It Works · Data refreshed daily, snapshot 2026-10-11.