Research from the lab
Long-form write-ups of our research on reasoning agents: the method, the numbers, and what we haven’t shown yet.
- GitOfThoughts: version-controlled reasoningWe saved an agent’s reasoning tree in an ordinary git repository, one commit per scored thought. Memory helped only when a stored trace was nearly a copy of the new problem, and git surfaced every contradictory lesson as a merge conflict.Pavan C Shekar
- Retrying isn’t the same as learning why you failedSampling again or asking a model to reflect tends to repeat the mistake. ReTreVal labels why a branch failed and retries from a different strategy: on MMLU-Pro, with no answer key in the loop, that added 10.8 points.Abhishek HS