Craftax framework evolves code policies via feedback cascade, reaching final stage 50% of time

Víctor Gallego · hf · 2026-11-17

A framework from Komorebi AI uses coding agents plus a feedback cascade to synthesize code-based policies in Craftax, a partially observed procedurally generated survival game with 100,000-action episodes. Scalar episode rewards stall optimization—final-level feedback arrives in under 1% of episodes—so the cascade shifts from episode returns to hierarchies of goals and belief updates scoreable densely from the simulator, letting the optimized policy reach the final stage 50% of the time (28% final-boss defeats). The paper flags a key trap: don't treat a subgoal's local success as the value source, use the continuation it leaves behind. The learned competence factors into a reusable goal compiler enabling zero-shot transfer, with zero inference-time model calls.

Original post →

More from coding & agent

coding & agent channel →