ASA: Training-Free Representation Engineering for Tool-Calling Agents
Youjin Wang, Run Zhou, Rong Fu, Shuaishuai Cao, Hongwei Zeng, Jiaxuan Lu, Sicheng Fan, Jiaqiao Zhao, Liangming Pan
2026-02-12
Summary
This paper addresses the challenge of getting large language model (LLM) agents to reliably use specific tools, especially when those tools change over time.
What's the problem?
Currently, making LLM agents good at using tools is tricky. You can try to carefully write instructions (prompt engineering) or define exactly how the tool should work (schema engineering), but these methods often break down when things change slightly. Another approach is to retrain the LLM, but that takes time, money, and can sometimes make the model forget things it already knew. The researchers found that even when the LLM *seems* to understand when a tool is needed based on its internal workings, it often hesitates to actually use it – there's a disconnect between what the model 'knows' and what it 'does'.
What's the solution?
The researchers developed a new method called Activation Steering Adapter (ASA). It's a small add-on that doesn't require any retraining of the LLM. Instead, it works by subtly adjusting the LLM’s internal thought process *while* it’s making a decision. It identifies which parts of the LLM’s internal calculations indicate a tool is needed and then boosts those signals, making the model more likely to use the tool. It uses a 'router' to choose the right adjustments for different tools and a 'gate' to make sure it's amplifying the *correct* reasons for using a tool, not just random noise.
Why it matters?
This work is important because it offers a way to make LLM agents much more reliable at using tools without the cost and complexity of constant retraining. It significantly improves the accuracy of tool use and reduces errors, all while being very small and easy to deploy. This could lead to more practical and useful AI assistants that can effectively interact with the real world through various tools and applications.
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.