Decision models can be better teachers than actors, a new study finds
Decision models can be better teachers than actors, a new study finds
By guthlabs.newsroom
A decision model does not need to be a strong agent to help train one. That is the central result of a November 12 preprint from researchers at Zhejiang University and Baidu, which compares choosing actions with judging completed trajectories. The distinction matters as builders explore cheaper, typed-output models for routing and control. Paper
The authors report that JEV-based supervision lifted a Qwen2.5-1.5B policy from 72.8% to 82.9% success on ALFWorld and from 56.8% to 71.4% on WebShop, relative to their GRPO baseline. Their experiments suggest that seeing an episode's outcome lets a judge recognize useful prerequisite steps that an acting model overlooks. These are author-reported benchmark results, not an independent production evaluation. Paper
The released implementation makes the boundary clearer: its teacher scores policy rollouts during training and is not called at test time. The researchers distinguish questions about eventual success from questions about an individual action's contribution. They also distinguish a partial trajectory from hindsight over the complete episode. The code repository is Apache-2.0 licensed. Implementation
Guth analysis: This is evidence for a training workflow, not a green light to let a confidence score authorize live actions. Retrospective credit assignment has information an online gate does not possess. An agent system should evaluate those roles separately, then keep permissions and human approvals outside the statistical score.
Limits and open questions: The result is a preprint on two environments. No independent reproduction was found in this pass. It does not establish equivalent gains for Gemini Decision 1, real-world purchasing, or high-stakes authorization. The useful next experiment is a held-out comparison of teacher questions and trajectory visibility before deploying any live gating policy.
Sources and citations
Revision history
Published .