Sign InOpen Brain
arXivPaperNeeds Review

Selective Agent Guidance via Entropy: Learning Autonomous Policies from Imperfect VLM Teachers

SAGE uses an expensive VLM only when a learner is uncertain, tests its advice against environment rewards, and distills useful guidance into a policy that needs no VLM at runtime.

arXiv · Sep 1, 2026
Open Source Open MarkdownOpen JSON
Source Summary

**SAGE** queries a VLM teacher only when the learner is uncertain, executes the proposed action during training, and distills it into a lightweight RL policy. Environment-derived advantages can downweight bad teacher advice.

Practical Implication

For visual or computer-use agents, consider the strong model a temporary exploration coach instead of the permanent policy. The learned agents needed **zero VLM calls at deployment** and sometimes exceeded their teachers on sparse-reward reasoning and navigation tasks.

Agent-Ready Context
**SAGE** queries a VLM teacher only when the learner is uncertain, executes the proposed action during training, and distills it into a lightweight RL policy. Environment-derived advantages can downweight bad teacher advice.

For visual or computer-use agents, consider the strong model a temporary exploration coach instead of the permanent policy. The learned agents needed **zero VLM calls at deployment** and sometimes exceeded their teachers on sparse-reward reasoning and navigation tasks.

Selective guidance helped when the teacher revealed high-reward trajectories. It added less when unguided exploration already worked or teacher actions produced little useful experience; the material does not quantify the call reduction or deployment savings.
Connected Context · Feed7 Judgment

This narrows strong-model guidance to a training-time exploration aid rather than a permanent computer-use dependency. Environment rewards can discount poor advice, and successful guidance can be distilled into a lightweight policy with no deployment-time VLM calls. The benefit is conditional, however: it depends on the teacher exposing rewarding trajectories that unguided exploration would not readily find.

TurnSight: Turn-Level Hindsight Self-Distillation for Tool-Integrated ReasoningBoth use execution-derived signals to improve tool-using policies: TurnSight assigns hindsight credit to individual turns, while SAGE uses environment advantages to determine whether uncertain-state teacher actions deserve reinforcement.Local Agentic Theory For Mobile Games — Shafik Quoraishee & Joanne Song, The New York TimesThe on-device agent work establishes memory, frame-time, and battery as deployment constraints; SAGE supplies a compatible training strategy by removing VLM calls from deployment, though it does not show those broader mobile budgets are met.The Dark Arts of Web Automation: Teaching Agents to Use Websites Like Humans — Corey Gallon, RexmoreBoth amortize costly exploration rather than repeatedly invoking general reasoning: the browser approach saves successful paths as deterministic code, while SAGE distills selectively guided experience into an autonomous learned policy.From RL to IRL — Gaurav Mishra, Amazon AGI LabSAGE addresses how a policy can learn useful actions despite imperfect guidance, but the runtime failure modes and handoff requirements identified here remain separate deployment safeguards that zero teacher calls do not resolve.
Context Map
agentimage#computer-use#agent-reliability#tool-use
Uncertainty
Selective guidance helped when the teacher revealed high-reward trajectories. It added less when unguided exploration already worked or teacher actions produced little useful experience; the material does not quantify the call reduction or deployment savings.