Research map / Large language models / Agents, code and math
Towards Execution-Grounded Automated AI Research
Chenglei Si, Zitong Yang, Yejin Choi, Emmanuel Candes, Diyi Yang, Tatsunori Hashimoto
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| use execution-guided evolutionary search instead of the GRPO baseline recipe | Qwen2.5-Math-1.5B post-training · GRPO environment | validation accuracy (%): 48.0% -> 69.4% | Improved |
| use execution-guided evolutionary search instead of the nanoGPT baseline | 124M GPT pre-training · nanoGPT environment | time to 3.28 val loss (min): 35.9 min -> 19.7 min | Improved |
| search for more epochs with Claude-4.5-Opus | execution-guided search · GRPO and nanoGPT environments | best performance per epoch | Improved |
| search for more epochs with Claude-4.5-Sonnet or GPT-5 | execution-guided search · GRPO and nanoGPT environments | best performance per epoch | No clear effect |
| execute Claude-4.5-Sonnet ideas with GPT-5 instead of self-execution | automated executor · GRPO environment | execution rate: 84% -> 42% | Hurt |
| execute Claude-4.5-Sonnet ideas with GPT-5 instead of self-execution | automated executor · nanoGPT environment | execution rate: 90% -> 78% | Hurt |
| train ideator with GRPO on execution reward (average reward) | Qwen3-30B-A3B · GRPO environment | average accuracy: 0.253 -> 0.343 | Improved |
| train ideator with GRPO on execution reward (average reward) | Qwen3-30B-A3B · nanoGPT environment | average validation loss: 5.150 -> 4.066 | Improved |
| train ideator with GRPO on execution reward (max reward) | Qwen3-30B-A3B · GRPO and nanoGPT environments | max reward per epoch | No clear effect |
| RL training for 68 epochs (diversity) | Qwen3-30B-A3B · nanoGPT environment | ideas that are RMSNorm swap or EMA (of 128): 51 -> 119 | Hurt |
| add dynamic prompt with previous trajectories to RL (ablation) | RL from execution reward · GRPO environment | reward | No clear effect |
| add length reward to RL (ablation) | RL from execution reward · GRPO environment | thinking length and total reward | Mixed |
Closest papers by meaning
- FML-bench: A Controlled Study of AI Research Agent Strategies from the Perspective of Search Dynamics (NeurIPS 2026 · similarity 0.88)
- MARS: Modular Agent with Reflective Search for Automated AI Research (ICML 2026 · similarity 0.87)
- From Reproduction to Replication: Evaluating Research Agents with Progressive Code Masking (ICLR 2026 · similarity 0.87)
- LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling (NeurIPS 2026 · similarity 0.86)
- InnovatorBench: Evaluating Agents’ Ability to Conduct Innovative AI Research (ICLR 2026 · similarity 0.86)
- Agentic AI Scientists Are Not Built For Autonomous Scientific Discovery (NeurIPS 2026 · similarity 0.86)
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.
