Chinese team open-sources ROME + ALE, a full-stack agentic training ecosystem
thisguyknowsai · x · 2026-10-06
A Chinese research team published a paper and open-sourced ROME + ALE (Agentic Learning Ecosystem), a full-stack recipe for training coding agents. The author's take is combative: most 'autonomous AI employee' demos are three ChatGPT calls wrapped in marketing, whereas ROME was built on real production-grade infrastructure — ROLL (RL framework), ROCK (sandboxed execution), iFlow CLI (orchestration). The 30B model (3B activated) scores 57.4% on SWE-bench Verified and ships with the contamination-free Terminal-Bench Pro benchmark.
More from coding & agent
- Her game hit 2M players and 100k+ concurrents in 4 days, yet people still call her a vibecoder — FrankFelixAI · 2026-10-06
- Creator shares 120-prompt workflow building an AI video game with Seedance 2.5, Godot and Magnific — techhalla · 2026-10-06
- Power users nuke all memory, workflows and setups whenever a new model ships — nateliason · 2026-10-06
- Cognition's SWE-2 outperforms Fable-5.1 and Opus 5.5 on Animation Bench, testers report — himanshustwts · 2026-10-06
- OpenAI makes Codex auto-review free, no longer counting against usage limits — mark_k · 2026-10-06
- DIY ESP32-S3 pocket voice assistant with e-paper display and Hermes agent routing — K8s-monk · 2026-10-06