ASA: Activation Steering for Tool-Calling Domain Adaptation

TL;DR

ASA addresses the domain adaptation challenge for tool-calling in large language models under frequent API churn by introducing a training-free Activation Steering Adapter (ASA). ASA performs a single-shot, mid-layer activation perturbation using a lightweight domain router, per-domain probes, and a mixtures-of-vectors (MoV) mechanism, gated by a probe-driven sign gate controlled by a single knob $\alpha$. It demonstrates that tool intent is linearly decodable from mid-layer activations yet requires selective, context-aware control to cross strict parser boundaries, enabling robust, cross-domain tool invocation without weight updates. Across multiple model scales and domains, ASA achieves LoRA-comparable gains with substantially lower overhead and strong transferability, providing a practical, deployment-friendly approach to managing dynamic tool ecosystems.

Abstract

For real-world deployment of general-purpose LLM agents, the core challenge is often not tool use itself, but efficient domain adaptation under rapidly evolving toolsets, APIs, and protocols. Repeated LoRA or SFT across domains incurs exponentially growing training and maintenance costs, while prompt or schema methods are brittle under distribution shift and complex interfaces. We propose \textbf{Activation Steering Adapter (ASA}), a lightweight, inference-time, training-free mechanism that reads routing signals from intermediate activations and uses an ultra-light router to produce adaptive control strengths for precise domain alignment. Across multiple model scales and domains, ASA achieves LoRA-comparable adaptation with substantially lower overhead and strong cross-model transferability, making it ideally practical for robust, scalable, and efficient multi-domain tool ecosystems with frequent interface churn dynamics.

Figures (9)

• Figure 1: Overview of tool-calling domain adaptation methods under rapidly evolving tool sets and interaction protocols, and our inference-time probe-gated activation control.
• Figure 2: ASA overview. We extract the last-token hidden state at layer $L$ during the pre-fill pass, route the input to a domain $\hat{d}$, estimate tool intent probability $p$, construct $\mathrm{MoV}=v_{\hat{d}}+\beta v_{\mathrm{global}}$, and inject $\Delta h=\mathrm{Gate}(h)\cdot \alpha \cdot \mathrm{MoV}$ once. No further interventions are applied during incremental decoding.
• Figure 3: Layer-wise probe sweep used to select the intervention depth.
• Figure 4: Combined geometry and causal diagnostic at the selected layer.
• Figure 5: Active steering performance on REST: domain-wise trends versus intervention strength $\alpha$.