Multi-Head Latent Control: A Unified Interface for LLM Agent Decision Making

Researchers introduced Multi-Head Latent Control, a lightweight layer that reads hidden-state traces from frozen LLMs and VLMs to predict agent actions.

It can decide when to defer to a stronger model, use tools, ask for clarification, answer directly, or abstain.

The method is trained post hoc, so the base model stays unchanged while routing and tool-use decisions improve across benchmarks.

This offers a cheaper, easier-to-maintain way to control agent behavior and cut expensive model calls without losing much performance.