Achieving SOTA Deep Research: Training Models Over Scaffolding

SimonShaoleiDu · x · 2026-07-30

In an interview, Simon Shaolei Du shared how they achieved SOTA performance in deep research by training the model itself rather than relying solely on scaffolding. He noted that the system runs up to 150 sub-agents and treats verification as a distinct layer rather than a final prompt. He also referenced findings suggesting that a single agent, even with a one-million-token context, struggles to carry hard research problems.

Related event: Simon Du Reveals SOTA Deep Research via Model Training(2 posts)→

Original post →

More from coding & agent

coding & agent channel →