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

2026-02-04

AI Summary

This work addresses the challenge of efficiently adapting large language models to diverse and dynamic toolsets and APIs across domains without requiring retraining. The authors propose a lightweight, training-free inference-time mechanism that introduces, for the first time, a routing strategy based on intermediate-layer activation signals. A hyper-lightweight router dynamically generates control strengths to achieve precise domain alignment. Remarkably, this approach matches the adaptation performance of LoRA across multiple model scales and domains—without any fine-tuning or prompt engineering—while substantially reducing computational overhead and demonstrating strong cross-model transferability.

Technology Category

• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Planning, Routing, and Scheduling: Planning with Language Models
• Natural Language Processing: (Large) Language Models

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

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.

Problem

• domain adaptation
• tool-calling
• LLM agents
• interface churn
• efficient adaptation

Innovation

• Activation Steering
• Tool-Calling
• Domain Adaptation
• Inference-Time Adaptation
• Training-Free