Zhejiang and Baidu authors say weak decision models can still teach strong RL policies
4.0/10Significance
A paper posted November 13 reports that as a teacher, JEV raises a Qwen2.5-1.5B policy from 72.8% to 82.9% success on ALFWorld. On WebShop it rises from 56.8% to 71.4%.
What changed
The authors say the same models, used as actors, are pulled away from a subgoal that must come first by an easier one. They report 22 to 30 USD of queries per training run.
Why it matters
It suggests a model that acts poorly can still be useful for training agents.
What remains uncertain
The results come from two benchmarks and small policy models.
What to watch
Watch for replication on other agent tasks.
Source
Why this ranks here
Clear numeric results on agent training, with a narrow scope.
- Impact4
- Reach4
- Novelty5
- Commitment4
- Coverage quality4
- Future potential5