ASA: Backbone-Training-Free Representation Engineering for Tool-Calling Agents
cs.SE, cs.AI
Submitted: 2026-02-04
Updated: 2026-08-28
Comments: Due to an unresolved dispute among the authors regarding the correctness of the experimental results and the validity of the conclusions, the team has decided to withdraw this paper until these issues can be fully resolved
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
The gist: Adapting LLM agents to domain-specific tool calling remains notably brittle under evolving interfaces.
Terminology
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.
Sources
- Phi-4-Mini Technical Report: Compact yet Powerful Multimodal Language Models via Mixture-of-LoRAs
- Therefore I am. I Think
- ReTool: Reinforcement Learning for Strategic Tool Use in LLMs
- Hammer: Robust Function-Calling for On-Device Language Models via Function Masking
- LLM Agents Already Know When to Call Tools -- Even Without Reasoning
- ToolAlpaca: Generalized Tool Learning for Language Models with 3000 Simulated Cases
- Steering Language Models With Activation Engineering
- PEFT A2Z: Parameter-Efficient Fine-Tuning Survey for Large Language and Vision Models
- Controlling Large Language Model Agents with Entropic Activation Steering
- Understanding Reasoning in Thinking Language Models via Steering Vectors
- Tool Calling is Linearly Readable and Steerable in Language Models
- Representation Engineering: A Top-Down Approach to AI Transparency
Related papers
- Falsification-Based Verification of LLM-Generated Optimization Models: Sound Test Batteries and Their Detection Limits
- GitSkills: A Dataset of Agent Skills on GitHub
- SABER: Benchmarking Operational Safety of LLM Coding Agents in Stateful Project Workspaces
- PackMonitor: Enabling Zero Package Hallucinations Through Decoding-Time Monitoring
- IntentCoding: Amplifying User Intent in Code Generation
- Incentives and Outcomes in Bug Bounties