Multi-Head Latent Control: A Unified Interface for LLM Agent Decision Making
A frozen model’s hidden states are used to decide when an agent should answer, call tools, defer, clarify, or abstain.
The paper introduces a lightweight control layer trained on latent traces from the same LLM or VLM backbone, without modifying that backbone. One head predicts whether to hand off to a stronger model; another chooses among clarification, tool use, abstention, or direct answering. In routed small-plus-large setups, it reports large-model usage reductions of up to 90.7% on AndroidWorld and 27-53% on average across benchmarks while keeping most large-model performance. The authors also report better tool-use decisions, including up to a 158% relative score gain and 65.5% fewer missed required tool calls. HF Daily Papers' note
The paper introduces a lightweight control layer trained on latent traces from the same LLM or VLM backbone, without modifying that backbone. One head predicts whether to hand off to a stronger model; another chooses among clarification, tool use, abstention, or direct answering. In routed small-plus-large setups, it reports large-model usage reductions of up to 90.7% on AndroidWorld and 27-53% on average across benchmarks while keeping most large-model performance. The authors also report better tool-use decisions, including up to a 158% relative score gain and 65.5% fewer missed required tool calls. HF Daily Papers' note
score 4