PMDC: leaderboard
Reward-model generalization by pairwise maximum discrepancy (PMDC, Luo et al., 2026): for every pair of reward models, the ten prompt-response pairs on which the two disagree most, drawn from 1,000 rewritten prompts with five LLM responses each, are judged by Claude Sonnet 4 in a forced choice, and the wins are fitted with a Bradley-Terry model. The score is the Bradley-Terry log-strength relative to ArmoRM-Llama3-8B-v0.1 (0); higher is better.
Metric: Bradley-Terry ranking score (log-strength; ArmoRM-Llama3-8B-v0.1 = 0) from Claude Sonnet 4 judgments on 450 maximum-discrepancy pairs. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Skywork-Reward-Gemma-2-27B | 2.49 | |
| 2 | QRM-Gemma-2-27B | 1.27 | |
| 3 | Reward-Model-Mistral-7B-instruct-unified | 0.75 | |
| 4 | URM-LLaMa-3.1-8B | 0.07 | #468 |
| 5 | ArmoRM-Llama3-8B-v0.1 | 0 | |
| 6 | Skywork-Reward-Llama-3.1-8B | -0.18 | |
| 7 | Skywork-Reward-V2-Qwen3-8B | -0.45 | |
| 8 | reward-model-deberta-v3-large-v2 | -0.51 | |
| 9 | Skywork-Reward-V2-Llama-3.2-3B | -1.01 | |
| 10 | Skywork-Reward-V2-Llama-3.1-8B | -1.06 |
Interactive version: theaggregate.ai/benchmark?slug=pmdc · How It Works · Data refreshed daily, snapshot 2026-11-12.