arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2602.04935cs.SEcs.AI
收藏

ASA:无需骨干训练的工具调用智能体表示工程

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

Renmin University of China(中国人民大学);University of Macau(澳门大学);Central South University(中南大学);Jiangxi Normal University(江西师范大学);Macau Millennium College(澳门 millennium 学院);Peking University(北京大学)

Youjin Wang, Run Zhou, Yingjie Ma, Rong Fu, Jiani Liang, Shuaishuai Cao, Min Huang, Tao Fang, Liangming Pan

AI总结:

针对大语言模型在工具调用中的惰性代理问题,提出一种无需训练、推理时激活干预的方法ASA,通过路由条件混合引导向量和探针引导门控,显著提升工具使用F1并降低误报率。

摘要:

将LLM智能体适应特定领域的工具调用在演化的接口下仍然明显脆弱。提示和模式工程易于部署,但在分布偏移和严格解析器下通常脆弱,而持续参数高效微调提高了可靠性,但代价是训练、维护和潜在的遗忘。我们识别出一个关键的惰性代理失败模式,其中工具必要性几乎可以从中间层激活完美解码,但模型在进入工具模式时仍然保守,揭示了表示-行为差距。我们提出激活引导适配器(ASA),一种无需训练、推理时控制器,执行单次中间层干预,并通过路由条件混合引导向量,结合探针引导符号门控来放大真实意图同时抑制虚假触发,从而定位工具领域。在MTU-Bench上使用Qwen2.5-1.5B,ASA将严格的工具使用F1从0.18提高到0.50,同时将假阳性率从0.15降低到0.05,仅使用约20KB的可移植资产且无需权重更新。

英文摘要:

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.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑