Multi-Head Latent Control uses hidden states to cut large-model calls by up to 90.7%
HuaweiTech · hf · 2026-07-27
Multi-Head Latent Control reads an LLM’s hidden states to steer agent decisions
The paper proposes Multi-Head Latent Control, a lightweight layer that sits on top of a frozen LLM/VLM and reads hidden-state trajectories to produce deployment-time control signals.
- Capability Head predicts whether the current model should solve the instance locally or defer to a stronger model.
- Resolution Head predicts whether to ask for more info, use a tool, abstain, or answer directly.
- The heads are trained only on latent traces from the same frozen backbone, so the approach can be added post hoc without changing the model.
Across language and vision-language settings, the method improves the quality-cost tradeoff of routed multi-model systems:
- In routed execution with small + large models, it cuts large-model usage by up to 90.7% on AndroidWorld and 27–53% on average across benchmarks, while keeping most of the large-model performance.
- It also improves tool-use decisions, with up to +158% relative score gain and 65.5% fewer missed-required tool calls.
More from coding & agent
- Applied AI is moving from chatbots to event-driven distributed systems — remilouf · 2026-07-27
- llmux manages vLLM and llama.cpp model swaps with one profile per model — Available-Message509 · 2026-07-27
- Windows DB-driven app builder asks what one AI subscription can really cover — Proud-Ticket-9379 · 2026-07-27
- Procedural and retrieval memory are becoming core agent primitives — goyalshaliniuk · 2026-07-27
- Graph memory and vector memory power modern AI retrieval stacks — goyalshaliniuk · 2026-07-27
- Episodic and semantic memory split experience from factual knowledge — goyalshaliniuk · 2026-07-27