AutoGym: Blueprint-First Generation of Verifiable Agent Gyms
- Type
- paper
- Venue
- arXiv
- Year
- 2026
- Source
- arxiv
- Access
- public
- Language
- en
- Added
- 2026-09-29
- Verified
- 2026-09-29
Summary
A framework that generates full agent environments — tasks, executable environments, and verifiers — from a minimal domain seed or from example trajectories, without hand-building each gym.
Keywords
agent environments · verifiable tasks · curriculum learning · synthetic tasks
Topics
agent environments, verifiable tasks, curriculum learning, synthetic tasks
Research notes
- Discovery: Shared in #random-papers as an arXiv link.
- Method: Blueprint-first generation: a compact domain blueprint is expanded into tasks, executable environments, and verifiers; explicit difficulty parameters plus performance-informed active curriculum synthesis control the training distribution.
- Detail from arXiv abstract only; code/dataset links not captured.