Plan-RewardBench - Single-Turn Planning (Hard): leaderboard
Metric: Pairwise accuracy (%) on the 158 hard single-turn planning pairs of the Plan-RewardBench chosen versus rejected agent trajectories (shared tool registry, multi-turn context, executed tool feedback), preference labels from a Gemini 3 Pro and GPT-5.1 judge panel with human audit; discriminative reward models score each trajectory, generative reward models and LLM judges choose between the pair; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 21 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | DeepSeek V3.2 Exp | 74.84 |
| 2 | Qwen 3 30B A3B 2507 Instruct | 74.68 |
| 3 | Qwen Plus | 74.68 |
| 4 | INF-ORM-Llama3.1-70B | 74.05 |
| 5 | InternLM2-7B-Reward | 74.03 |
| 6 | Qwen 3 4B 2507 Instruct | 73.73 |
| 7 | DeepSeek R1 | 73.58 |
| 8 | FsfairX-LLaMA3-RM-v0.1 | 72.78 |
| 9 | Skywork-Reward-V2-Llama-3.1-8B | 72.15 |
| 10 | Qwen 3 235B A22B 2507 Instruct | 71.36 |
| 11 | Skywork-Reward-V2-Qwen3-8B | 70.57 |
| 12 | Qwen Max | 68.99 |
| 13 | Gemini 3 Flash (Preview) | 67.25 |
| 14 | GPT-5 | 62.18 |
| 15 | Qwen 3 235B A22B 2507 (Thinking) | 60.13 |
Interactive version: theaggregate.ai/benchmark?slug=plan-rewardbench-single-turn-planning-hard · How It Works · Data refreshed daily, snapshot 2026-10-07.