MLS-Bench (Agent) - Reinforcement Learning: leaderboard
MLS-Bench (Agent) Reinforcement Learning scores the final submission after up to 20 actions and three test runs on the 13 Reinforcement Learning tasks of MLS-Bench (2026), where a model must invent a better ML method inside a frozen research codebase. Task scores are normalized so the worst reproduced human baseline is 0 and the best is 50, on a 0 to 100 scale where higher is better.
Metric: 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 13 MLS-Bench tasks of the Reinforcement Learning area; 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 February 2028. 5 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Claude Opus 4.6 | Thinking | 28.2 | #114 |
| 2 | Gemini 3.1 Pro (Preview) | High | 26.9 | #102 |
| 3 | Qwen 3.6 Plus | Thinking | 13.1 | #245 |
| 4 | DeepSeek V3.2 | Thinking | 12.4 | #425 |
| 5 | GPT-5.4 | High | 12.4 | #138 |
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-reinforcement-learning · How It Works · Data refreshed daily, snapshot 2026-11-22.