CS329A Self-Improving AI Agents, Part 7: Self-Improvement and Deep Research Agents

Name: CS329A Self-Improving AI Agents, Part 7: Self-Improvement and Deep Research Agents
Uploaded: 2026-09-28T00:00:00.000Z
Duration: 4346 s
Channel: Stanford Online
Description: AlphaCode reached competitive-programming results via huge sample budgets and clustering; AlphaCode2 fine-tuned Gemini Pro with a learned scoring model to reach the 85th percentile. Search-O1 applies the same lesson to research: retrieve at uncertainty points and reason over documents.

Stanford CS329A Self-Improving AI Agents | Part 7 | Self-Improvement and Deep Research Agents

SOStanford Online@stanfordonline

Full transcript

English

Summary：AlphaCode reached competitive-programming results via huge sample budgets and clustering; AlphaCode2 fine-tuned Gemini Pro with a learned scoring model to reach the 85th percentile. Search-O1 applies the same lesson to research: retrieve at uncertainty points and reason over documents.

Watch on YouTube

Core points (3)

1Solve rate scales with sample budget, but selection and clustering become the bottleneck.
2Search-O1 fires retrieval only when the reasoner signals uncertainty.
3Prompted search (Search-O1) and RL-trained search (Search-R1) trade flexibility for stability.