MLS-Bench (Agent): leaderboard

MLS-Bench (Agent) scores the final submission after up to 20 actions and three test runs when a model must invent a better ML method inside a frozen research codebase, over 140 tasks in 12 research areas of MLS-Bench (2026). Each task score is normalized so the worst reproduced human baseline is 0 and the best is 50, on a 0 to 100 scale where higher is better; the board averages the 12 area scores.

Metric: Mean of the 12 area scores, each the mean normalized task score (0-100): each metric is anchored so the worst reproduced human baseline scores 0 and the best 50 (a metric's theoretical optimum maps to 100), metrics are averaged with human weights within a setting and by geometric mean across at least three generalization settings, over the 140 method-invention tasks of MLS-Bench (the full set, not MLS-Bench-Lite); Agent protocol: the model's final submission after at most 20 actions including 3 test calls; high reasoning effort with a 10,000-token thinking budget, web search disabled, fixed seeds; higher is better. Source: arxiv.org. Saturation forecast: Around May 2028. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Opus 4.6Thinking36#114
2Gemini 3.1 Pro (Preview)High34.9#102
3GPT-5.4High27.8#138
4DeepSeek V3.2Thinking26.3#425
5Qwen 3.6 PlusThinking23.2#245

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=mls-bench-agent · How It Works · Data refreshed daily, snapshot 2026-11-22.