Themis-CodeRewardBench - Security Hardness: leaderboard
The security split of Themis-CodeRewardBench (Paul, Glavaš and Gurevych, 2026): 991 code preference pairs from vulnerability-fix commits, CodePrefBench, Vul4J, SecBench and NoFunEval, where the preferred response is the hardened one. Preference accuracy of pointwise reward-model scores in percent; higher is better.
Metric: Preference accuracy (%; share of the benchmark's preference pairs in which the reward model scores the preferred code response above the rejected one, pointwise and reference-free; Security Hardness criterion, 991 pairs: vulnerability-fix commits, CodePrefBench Security, Vul4J, SecBench and NoFunEval Security). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 51 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Themis-RM-14B | 95.36 | |
| 2 | Themis-RM-32B | 94.55 | |
| 3 | Themis-RM-8B | 93.34 | |
| 4 | Themis-RM-4B | 92.94 | |
| 5 | Themis-RM-1.7B | 89 | |
| 6 | Themis-RM-0.6B | 87.69 | |
| 7 | Athene-RM-70B | 78.3 | |
| 8 | Skywork-Reward-V2-Qwen3-8B | 74.97 | |
| 9 | Athene-RM-8B | 74.77 | |
| 10 | Llama-3.1-Tulu-3-70B-SFT-RM-RB2 | 73.76 |
Interactive version: theaggregate.ai/benchmark?slug=themis-coderewardbench-security-hardness · How It Works · Data refreshed daily, snapshot 2026-11-12.