Themis-CodeRewardBench - Readability and Maintainability: leaderboard
The readability and maintainability split of Themis-CodeRewardBench (Paul, Glavaš and Gurevych, 2026): 1,499 code preference pairs from code-style commits and NoFunEval maintainability fixes, where the preferred response is the cleaner 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; Readability and Maintainability criterion, 1,499 pairs: code-style commits and NoFunEval Maintain). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 51 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Themis-RM-32B | 87.59 | |
| 2 | Themis-RM-14B | 87.32 | |
| 3 | Themis-RM-8B | 86.06 | |
| 4 | Themis-RM-4B | 84.52 | |
| 5 | Themis-RM-1.7B | 80.05 | |
| 6 | Themis-RM-0.6B | 78.79 | |
| 7 | Athene-RM-70B | 75.45 | |
| 8 | Skywork-Reward-V2-Qwen3-8B | 75.05 | |
| 9 | Skywork-Reward-V2-Qwen3-4B | 74.38 | |
| 10 | Llama-3.1-Tulu-3-70B-SFT-RM-RB2 | 73.58 |
Interactive version: theaggregate.ai/benchmark?slug=themis-coderewardbench-readability-and-maintainability · How It Works · Data refreshed daily, snapshot 2026-11-12.