A Model That Fails 42.9% of Tasks Can Still Teach an Agent to 82.9%

rohanpaul_ai · x · 2026-11-15

A Zhejiang University paper shows an AI model bad at a task can still train another model to do it well. The decision model JEV completes only 42.9% of ALFWorld household tasks itself, but as a step-level teacher reviewing another AI's moves, it lifts a Qwen2.5-1.5B policy from 72.8% to 82.9% success versus GRPO outcome rewards alone.

Key points:

Agent RL usually rewards only final outcomes, giving every step equal credit; per-step scoring normally needs environment labels or an LLM judge at every action.
As judges seeing whole trajectories, models correctly credit steps toward essential subgoals; asking whether each action helped—rather than whether the task will succeed—separates useful from useless actions.
As actors, all six tested models get pulled off a required subgoal; as judges, the same models handle it fine.

Advice: pick step graders by how well they judge recorded runs, not how well they act. Full paper details covered in a separate post.

Original post →

A Model That Fails 42.9% of Tasks Can Still Teach an Agent to 82.9%

More from coding & agent