ECHO Paper Lands NeurIPS Spotlight: Terminal Agents Learn World Models for Free, Doubling GRPO Pass@1
DimitrisPapail · x · 2026-09-25
ECHO: Terminal Agents Learn World Models for Free
The paper "ECHO: Terminal Agents Learn World Models for Free" (Vaishnavi Shrivastava, Dimitris Papailiopoulos et al.) has been accepted as a NeurIPS 2026 Spotlight.
- Insight: CLI agents are the closest thing LLMs have to an embodied setting—the terminal's response stream (stdout, errors, logs) is a supervision signal that standard GRPO-style RL throws away, leaving failed rollouts with almost no gradient signal.
- Method: ECHO adds an auxiliary loss on top of the policy-gradient loss, training the policy to predict environment observation tokens resulting from its own actions. It reuses the same forward pass as GRPO, needs no extra rollouts, and turns terminal feedback into dense supervision.
- Results: Pass@1 on TerminalBench-2.0 roughly doubles—Qwen3-8B improves from 2.70% to 5.17%, and Qwen3-14B from 5.17% to 10.79%. ECHO policies also predict terminal dynamics far better on held-out rollouts, and from base Qwen3-8B it matches expert-SFT-then-GRPO performance without any expert demonstrations.
- Paper and code are both available.
More from coding & agent
- Eleanor Berger's Free Lesson: Five Common Mistakes with Agentic Factories — intellectronica · 2026-09-25
- aeon: a language that uses types and logic to guardrail AI agents from nonsensical plans — dscape · 2026-09-25
- Indie devs use the Cloudflare API to auto-provision per-customer SaaS subdomains — gregmushen · 2026-09-25
- Only Then Labs launches ProofPress to carry research evidence across AI agents — tallmetommy · 2026-09-25
- Dev open-sources Jev reasoning lab: model routing and adversarial peer-review experiments — arthurcolle · 2026-09-25
- mitsuhiko complains Opus 5.5 keeps editing files via bash, mulls a blocking extension — mitsuhiko · 2026-09-25