Research map / Reinforcement learning / Visual and representation RL
Reinforcement Learning for Machine Learning Engineering Agents
Sherry Yang, Joy He-Yueya, Percy Liang
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| RL training of Qwen2.5-3B instead of prompting it with AIDE | Qwen2.5-3B · MLE-Bench: random-acts-of-pizza | task score: 0.589 -> 0.663 | Improved |
| RL training of Qwen2.5-3B instead of prompting it with AIDE | Qwen2.5-3B · MLE-Bench: tweet-sentiment-extraction | task score: 0.027 -> 0.596 | Improved |
| RL-trained Qwen2.5-3B instead of prompting Claude-3.5-Sonnet | Qwen2.5-3B · MLE-Bench: leaf-classification | log loss: 0.436 -> 0.124 | Improved |
| RL-trained Qwen2.5-3B instead of prompting Claude-3.5-Sonnet | Qwen2.5-3B · MLE-Bench: us-patent-phrase-to-phrase-matching | task score: 0.805 -> 0.527 | Hurt |
| RL-trained Qwen2.5-3B instead of prompting Claude-3.5-Sonnet | Qwen2.5-3B · MLE-Bench (12 tasks average) | relative improvement: 22% | Improved |
| RL-trained Qwen2.5-3B instead of best GPT-4o scaffold | Qwen2.5-3B · MLE-Bench (12 tasks average) | relative improvement: 17.7% | Improved |
| run GPT-4o AIDE for 100 hours instead of 24 | GPT-4o · MLE-Bench | task score | No clear effect |
| add duration-aware gradient weighting (ablation) | RL training · random-acts-of-pizza | execution time of solutions | Improved |
| add environment instrumentation partial credit (ablation) | RL training · MLE-Bench: plant-pathology-2020-fgvc7 | average task score | Improved |
| add explicit self-improvement prompt (ablation) | Qwen2.5-3B RL · MLE-Bench (12 tasks average) | relative improvement: 8% | Improved |
| add improve-previous-solution prompt (ablation) | Qwen2.5-3B RL · MLE-Bench: leaf-classification | log loss: 0.469 -> 0.124 | Improved |
| add improve-previous-solution prompt (ablation) | Qwen2.5-3B RL · MLE-Bench: plant-pathology-2020-fgvc7 | task score: 0.949 -> 0.970 | Improved |
Closest papers by meaning
- ML-Agent: Reinforcing LLM Agents for Autonomous Machine Learning Engineering (ICML 2026 · similarity 0.91)
- MLE-Dojo: Interactive Environments for Empowering LLM Agents in Machine Learning Engineering (NeurIPS 2025 · similarity 0.91)
- MLE-STAR: Machine Learning Engineering Agent via Search and Targeted Refinement (NeurIPS 2025 · similarity 0.89)
- MLE-Smith: Scaling MLE Tasks with Automated Multi-agent Pipeline (ICLR 2026 · similarity 0.88)
- MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering (ICLR 2025 · similarity 0.88)
- AI Research Agents for Machine Learning: Search, Exploration, and Generalization in MLE-bench (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
