analiza aiqbench
ASA introduces a training-free, inference-time controller that bridges the representation-behavior gap in LLM agents by steering mid-layer activations to improve strict tool-use reliability without weight updates. This approach significantly boosts F1 scores on MTU-Bench while drastically reducing false positives, offering a lightweight alternative to fine-tuning for evolving tool interfaces. The method's portability, requiring only ~20KB of assets, makes it a practical solution for maintaining agent robustness against distribution shifts.
ASA: Backbone-Training-Free Representation Engineering for Tool-Calling Agents
arXiv:2602.04935v3 Announce Type: replace-cross 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.
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.
The manuscript consists of 24 pages formatted in the ACL style. Youjin Wang, Run Zhou, and Yingjie Ma contributed equally to this work. Tao Fang and Liangming Pan are the co-corresponding authors
Software Engineering (cs.SE); Artificial Intelligence (cs.AI)
arXiv:2602.04935 [cs.SE]
(or arXiv:2602.04935v3 [cs.SE] for this version)
Text integral sub licența deschisă a publicației · 196 cuvinte
Subiecte conexe
[AINews] High Taste Testers Ep. 1
Latent Space · 17 oct. 2026
ASA: Backbone-Training-Free Representation Engineering for Tool-Calling Agents
arXiv · 1 sept. 2026
[AINews] Meta Muse Wonder & Wonder Full: MSL’s Claude 5.5-tier Frontier Model
Latent Space · 13 oct. 2026
News
Quantum Computing Report · 25 iul. 2026