ASA: Training-Free Representation Engineering for Tool-Calling Agents

TL;DR

This paper introduces ASA, a training-free method that improves tool-calling accuracy in language models by intervening at mid-layer activations, significantly enhancing reliability without additional training.

Contribution

We propose Activation Steering Adapter (ASA), a novel inference-time intervention technique that bridges the representation-behavior gap for tool calling in language models without training.

Findings

ASA increases strict tool-use F1 from 0.18 to 0.50.

ASA reduces false positive rate from 0.15 to 0.05.

Requires only 20KB of assets and no model fine-tuning.

Abstract

Adapting LLM agents to domain-specific tool calling remains notably brittle under evolving interfaces. Prompt and schema engineering is easy to deploy but often fragile under distribution shift and strict parsers, while continual parameter-efficient fine-tuning improves reliability at the cost of training, maintenance, and potential forgetting. We identify a critical Lazy Agent failure mode where tool necessity is nearly perfectly decodable from mid-layer activations, yet the model remains conservative in entering tool mode, revealing a representation-behavior gap. We propose Activation Steering Adapter (ASA), a training-free, inference-time controller that performs a single-shot mid-layer intervention and targets tool domains via a router-conditioned mixture of steering vectors with a probe-guided signed gate to amplify true intent while suppressing spurious triggers. On MTU-Bench with Qwen2.5-1.5B, ASA improves strict tool-use F1 from 0.18 to 0.50 while reducing the false positive rate from 0.15 to 0.05, using only about 20KB of portable assets and no weight updates.