Plan-RewardBench - Error Recovery: leaderboard

Plan-RewardBench (2026) tests reward models and LLM judges on preferring the better of two tool-using agent trajectories, with preference labels from a judge panel audited by humans. This board is pairwise accuracy in percent on the pairs that test recovery from tool errors; higher is better.

Metric: Pairwise accuracy (%) on the 361 robust error-recovery 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. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Skywork-Reward-V2-Qwen3-8B75.35
2Qwen 3 30B A3B 2507 Instruct74.93#1100
3Qwen Plus73.75#820
4INF-ORM-Llama3.1-70B69.78
5GPT-569.39#178
6Qwen 3 235B A22B 2507 Instruct68.84#616
7Qwen 3 4B 2507 Instruct68.7#2221
8Qwen Max68.63#1092
9DeepSeek R168.01#482
10Gemini 3 Flash (Preview)67.31#152
11RRM-32B67.15#1115
12InternLM2-7B-Reward67.05
13DeepSeek V3.2 Exp66.76#439
14Qwen 3 235B A22B 2507 (Thinking)66.48#466
15Skywork-Reward-V2-Llama-3.1-8B65.1

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=plan-rewardbench-error-recovery · How It Works · Data refreshed daily, snapshot 2026-11-21.