AgentGarten trains agents in rendered code worlds, learning in 4 rounds vs millions for RL
MirroS-Lab · hf · 2026-10-09
MirroS-Lab released AgentGarten on Hugging Face, a framework for training agents through exploration of interactive virtual worlds.
Key points:
- Couples simulators/game engines with a shared neural renderer: backends maintain persistent world state and run program-defined interaction rules, while the renderer generates realistic visual observations from structured conditions.
- The renderer adapts a pretrained video model and distills it via Adversarial Forcing, which makes history prefilling differentiable through exact replay and adds real-data adversarial supervision.
- Agents distill each round of experience into playbooks that later agents inherit and refine.
Experiments show large learning-efficiency gains: agents learn from just 4 rounds versus millions for a conventional RL counterpart. New worlds can be written as code, letting environments scale in number and difficulty alongside agents.
Related event: AgentGarten: Agents Learn in Neural-Rendered Interactive Worlds(2 posts)→
More from coding & agent
- Cresta hosts Anthropic-sponsored voice AI agent hackathon at SF Tech Week — schwentker · 2026-10-09
- Building a sellable 3D wuxia ARPG with AI: Tripo modeling plus Claude coding, pitfalls included — tinyfool · 2026-10-09
- hashcards v0.5.0 ships a web interface for browsing flashcard collections — zetalyrae · 2026-10-09
- Dev Vibecodes a Counter-Strike Remake in a Week, Runs Smoothly in Any Browser — TAbrodi · 2026-10-09
- Claude Code auto-compacts while idle, before your prompt cache expires to save usage limits — JeremyNguyenPhD · 2026-10-09
- Hermes Agent Adds Proactive Smart UI in Hermes Desktop, Rendering Embeds Unprompted — Teknium · 2026-10-09