Research question: separating failure history from a difficult game state

[GBX-Max1220]

Hi Neel and team, I’m Max. I emailed Neel earlier this month about learning from failures in FLE.

The observation that agents can perform worse after repeated failures made me curious about two possible causes: the history they see, and the difficult game state they have reached.

Would it be useful to separate those effects by starting from the same saved state and varying the supplied history—for example, the full failed attempts, a factual summary, or a summary containing only verified lessons? Any comparison would need to preserve the information required to understand the current state.

I’d be interested in helping design and analyze a small study if this fits your research plans. My recent work involves agent evaluation and failure attribution, and I can contribute remotely from China.