Heading-Specific Activation Steering: Controlling LLM Tool Use

Controlling Tool Use with Heading-Specific Activation Steering, a paper by Yuqi Chen, Vincent Siu, Yang Liu, Dawn Song and Chenguang Wang submitted to arXiv on 7 Jul 2026, demonstrates that internal activation steering can causally change whether large language models call external tools. The authors extract steering vectors from heading-anchor positions and show those vectors exert bidirectional causal control over tool-invocation behavior across five open-source models and three domains.

What did the paper show?

The paper shows that steering vectors derived from heading-anchors can both suppress and encourage tool invocation, succeeding across five open-source models and three domains, and doing best where parametric reasoning suffices. The authors report that steering vectors suppress unnecessary tool use most effectively in tasks that do not require external tools, and that the vectors have bidirectional causal effects on the model’s decision to invoke a tool.

They also report geometric irregularities in the internal signatures tied to tool use. Geometric analysis in the paper finds that tool-invocation steps exhibit "diffuse, bimodal alignment" with the suppression vector instead of a simple linear negative alignment. Different tool types recruit largely distinct internal signatures, producing low cross-tool feature overlap.

How does heading-specific activation steering work?

Heading-specific activation steering uses steering vectors extracted from positions tied to headings in the input, then inserts or manipulates those vectors during inference to change the model’s tool-use decisions. The paper locates heading-anchor positions in context, extracts steering vectors from those positions, and applies them to control whether the model invokes an external tool at subsequent steps.

The method treats tools as non-parametric context items that lack a fixed encoding in model weights. Because tools are present only in context at inference time, their influence must be captured from activations, and the heading-anchored vectors act as levers applied at runtime. The authors show these levers work across multiple open-source models and domains, demonstrating causal control rather than mere correlation.

What are the technical observations about internal geometry?

The steering vectors produce measurable causal effects but do not map to a clean linear internal code for tool use. Geometric analysis in the paper reveals that alignment between tool-invocation steps and the suppression vector is diffuse and bimodal rather than uniformly negative, and that signatures differ across tool types with little overlap. The authors contrast these properties with vectors extracted for parametrically grounded concepts, implying tool-related activations behave differently because tools are non-parametric context items.

Those geometric features leave an open question connecting irregular geometry to causal effectiveness. The paper states that the relationship between geometric irregularity and observed causal control remains unresolved.

Why it matters

If tool-use decisions can be controlled from activations, model developers gain a runtime mechanism to reduce unnecessary tool calls without changing model weights or tool implementations. That can lower latency, reduce external API usage, and make model behavior more predictable when tools are not needed. The paper’s finding that suppression works best where parametric reasoning suffices suggests a practical division: apply activation steering to avoid tool overuse in tasks the base model can already handle.

The geometric findings also matter for safety and interpretability. Distinct, low-overlap signatures for different tool types complicate any one-size-fits-all steering approach and suggest tool-specific controls will be necessary.

What to watch

Look for follow-up work connecting the paper’s geometric observations to mechanism-level explanations and for experiments extending the technique to additional models and tool families. The authors flag the relationship between the irregular internal geometry and causal effectiveness as an open question, making that a clear next milestone to confirm whether steering can be generalized or must be bespoke per tool type.

For reference, the paper is available on arXiv as arXiv:2607.05790 and carries the submission date 7 Jul 2026.