Themis-CodeRewardBench - Memory Efficiency: leaderboard
The memory-efficiency split of Themis-CodeRewardBench (Paul, Glavaš and Gurevych, 2026): 289 code preference pairs from memory-improving commits and NoFunEval, where the preferred response uses less memory. 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; Memory Efficiency criterion, 289 pairs: memory-improving commits and NoFunEval Memory). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 51 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Themis-RM-32B | 95.16 | |
| 2 | Themis-RM-14B | 94.46 | |
| 3 | Themis-RM-4B | 92.39 | |
| 4 | Themis-RM-8B | 92.04 | |
| 5 | Themis-RM-1.7B | 86.85 | |
| 6 | Themis-RM-0.6B | 86.85 | |
| 7 | Athene-RM-70B | 77.16 | |
| 8 | Skywork-Reward-V2-Qwen3-4B | 72.66 | |
| 9 | CodeScaler-8B | 72.32 | |
| 10 | Skywork-Reward-V2-Qwen3-8B | 71.63 |
Interactive version: theaggregate.ai/benchmark?slug=themis-coderewardbench-memory-efficiency · How It Works · Data refreshed daily, snapshot 2026-11-12.