Benchmarks
357 sources · 1,263 models × 3,079 benchmarks · 162,127 scoresPeking University DAGroup's RSI-Infra Bench Preview asks whether a coding agent can make real training jobs from public repositories run faster in 10 hours without changing the training. Its eight tasks (LLaVA, YOCO, NanoChat, BAGEL SFT, R3, 3D Gaussian splatting, and Slime GRPO and on-policy distillation) run on one to four Hopper GPUs; a frozen judge times the original and the agent's code back to back and accepts only bit-identical results or results within 3 sigma of the original's seed-to-seed variation, at no lower precision, and a rejected or slower run counts as 1x. Each task's speedup is normalized as ln(speedup) / ln(best speedup on the task), and an Epoch Capabilities Index-style fit over all models and tasks gives the score, scaled so that Claude Opus 5.5 = 40 and Kimi K3.1 Pro = 10. Each model joining the board refits the scale and can move every other score, so the score is relative to the board's cohort and this is a diagnostic.
No signal or difficulty estimate. Held out of the fit by design.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 | not reported · Claude Code | 40.0 | |
| 2 | Claude Sonnet 5.5 | not reported · Claude Code | 36.7 | |
| 3 | GPT-6.2 Astra | max · Codex | 28.4 | |
| 4 | GPT-6.2 Sol | max · Codex | 21.8 | |
| 5 | MiniMax M3.1 Flash | not reported · Codex | 20.6 | |
| 6 | Claude Haiku 5.5 | not reported · Claude Code | 19.3 | |
| 7 | DeepSeek V4.1 Flash 1017 | not reported · DeepSeek Harness | 14.2 | |
| 8 | GLM-5.3 | not reported · Claude Code | 13.7 | |
| 9 | Kimi K3.1 Pro | not reported · Claude Code | 10.0 |