Amazon AGI's AutoGym auto-generates tasks, environments, and verifiers for agent RL training
omarsar0 · x · 2026-09-29
A new Amazon AGI paper introduces AutoGym, a framework that auto-generates complete RL gyms — task, executable environment, and verifier — from a small domain seed or past model trajectories.
Three key mechanisms:
- Blueprint-first generation: defines the valid solution space, environment requirements, and verification criteria before materialization, making solvability a construction prerequisite rather than a post-hoc check
- Explicit generation parameters: fine-grained control over task topology, interaction depth, capability axes, obfuscation, and distractor composition for precise difficulty steering
- Active curriculum synthesis: performance-informed calibration adjusts parameter distributions as model capabilities evolve
The paper argues static task sets saturate and get contaminated, single-pass synthesis yields cosmetically hard tasks, and LLM judges are unreliable. Results span productivity and temporal-reasoning settings.
More from Research
- Aether AI releases 16B CausalWM, a world model that reasons about causality before generating future frames — jiqizhixin · 2026-09-29
- Calibrated Importance Sampling Fixes Training-Inference Mismatch in LLM RLVR — Tianrun Yu · 2026-09-29
- ZJU Releases EMem-Bench: 2,554 Episodes to Test Embodied Agent Memory — OmniAI-ZJU · 2026-09-29
- Tencent's AdaTutoRank Trains Setwise RAG Rerankers via Adaptive Tutoring Optimization — tencent · 2026-09-29
- Stanford lands $25M NIH grant to build AI center for personalized dementia care — StanfordAILab · 2026-09-29
- RL Gains May Hit a Wall: Verifiable Data Limits and Stalling nanochart Benchmarks — QuintinPope5 · 2026-09-29