Plan-RewardBench - Safety Refusal: 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 pairs that test appropriate safety refusal, including avoiding over-refusal on benign requests; higher is better.

Metric: Pairwise accuracy (%) on the 51 safety-refusal 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: Around February 2027. 21 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-584.8#178
2Qwen 3 235B A22B 2507 (Thinking)78.92#466
3Kimi K2 (Thinking)78.63
4Gemini 3 Flash (Preview)78.43#152
5DeepSeek V3.2 Exp75#439
6DeepSeek R172.55#482
7Qwen Max71.08#1092
8Qwen 3 235B A22B 2507 Instruct65.69#616
9M-Prometheus-14B62.44#1255
10RRM-32B60.3#1115
11INF-ORM-Llama3.1-70B58.53
12Qwen 3 4B 2507 Instruct57.35#2221
13Skywork-Reward-V2-Qwen3-8B56.86
14RM-R1-DS-Distill-Qwen-32B56.62
15Qwen Plus55.88#820

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-safety-refusal · How It Works · Data refreshed daily, snapshot 2026-11-21.