Decision models act poorly but teach well: GRPO teacher lifts Qwen 1.5B to 82.9% on ALFWorld
omarsar0 · x · 2026-11-13
A Zhejiang University/Baidu paper compares decision models as actors, judges, and teachers for long-horizon agents. All six models tested (three decision models, three general LLMs) pulled away from prerequisite subgoals as actors on ALFWorld, yet as judges seeing full trajectories they credited the steps they skipped. Used as a teacher scoring each action for GRPO, one decision model raised a Qwen2.5-1.5B policy from 72.8% to 82.9% on ALFWorld and from 56.8% to 71.4% on WebShop — above the 42.9% the model achieves as an actor itself. A poor actor can teach well, if asked the right question. Code is open-sourced.
Related event: Zhejiang/Baidu Study: Decision Models Teach Better Than They Act(2 posts)→
More from coding & agent
- "Code review is dead": nobody actually reads an 8,294-line agent diff — prasenx · 2026-11-13
- AI educator alfcnz: students lack coding-agent skills, so he'll teach a method — alfcnz · 2026-11-13
- Decision models for agent tool selection: 50 tools to 3, 47% fewer tokens, 22% faster — rajistics · 2026-11-13
- Anthropic engineer: Claude iterated meeting calendar designs live in one brainstorm — Flomerboy · 2026-11-13
- Nuance Live preview: build real-time audio-video apps where an AI persona calls your tools — koltregaskes · 2026-11-13
- Agent Deleted 6 Lines, Caught 3 Weeks Later: 400 Customers Wrongly Charged — Imaginary-Mix-7777 · 2026-11-13