[论文] One Task, Many Interfaces: Simple Cross-Interface Consistency Training...

论文概要

研究领域: NLP
作者: Artin Tajdini, Manpreet Kaur, Peng Qi
发布时间: 2026-10-14
arXiv: 2610.20576

中文摘要

后训练将 LLM 代理的能力绑定到其训练所用的界面上。对界面进行语义保持的扰动——例如重命名动作或更改符号表示——底层任务并未改变，却会让代理的表现变差。对每条样本独立评分意味着界面随机化仅将这种依赖从单一界面扩展到训练界面分布，分布外扰动仍会削弱代理能力。我们提出跨界面策略正则化（CIPR），一种自监督目标：将动作名替换为无信息标识符，并在每个训练样本下用两种构造上等价动作天然对齐的界面呈现，惩罚两种界面诱导的动作分布之间的 KL 散度，使策略跨界面保持一致。我们在有监督后训练和强化学习后训练中评估 CIPR，均改善了鲁棒性。在代理式函数调用任务上，留出界面扰动使后训练代理的准确率下降近 25%，CIPR 恢复了 10.5%，同时在未变化界面上的准确率不受影响。

原文摘要

Post-training binds the competence of an LLM agent to the interface it was trained on. Semantics-preserving perturbations of that interface, such as renaming an action or changing the notation, leave the underlying task unchanged yet make the agent worse at it. Scoring each sample independently means interface randomization only broadens this dependence from a single interface to the training interface distribution, so out-of-distribution perturbations still degrade the agent. We propose Cross-Interface Policy Regularization (CIPR), a self-supervised objective that replaces action names with uninformative identifiers and presents each training sample under two interfaces in which equivalent actions are identified by construction. CIPR penalizes the KL divergence between the action distribu...

自动采集于 2026-10-16

#论文 #arXiv #NLP #小凯