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.
More from coding & agent
- OutfitterBench tests LLM agents on 200-day rental business, exposing commitment failures — Jie Huang · 2026-11-17
- Entity disambiguation: sentence context hits 97.9% vs 63.5% for name matching — YvesMulkers · 2026-11-17
- XState v6 Beta Launches: State Machines Rewritten for the AI Agent Era — mattpocockuk · 2026-11-17
- Matt Pocock shares a prompt for having AI agents evaluate new tech via draft PRs — mattpocockuk · 2026-11-17
- Motion launches Claude Code for Video Editing: raw footage plus one prompt, auto-edited — CurieuxExplorer · 2026-11-17
- Watch the replay: Lauren Tan's AI agent workflow shipping 3,000 PRs/month — FinanceYF5 · 2026-11-17