arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2603.20899cs.CLcs.AI
收藏

缓解语言模型中的捷径推理:一种梯度感知的训练方法

Mitigating Shortcut Reasoning in Language Models: A Gradient-Aware Training Approach

Arizona State University(亚利桑那州立大学);Clemson University(克莱姆森大学);University of Kansas(堪萨斯大学)

Hongyu Cao, Kunpeng Liu, Dongjie Wang, Yanjie Fu

AI总结:

本文提出SART框架,通过梯度感知技术检测并缓解语言模型中的捷径推理,实验显示在受控推理基准上提升准确率和鲁棒性。

摘要:

大型语言模型表现出强大的推理能力,但往往依赖于表面模式匹配和答案记忆等捷径而非真正的逻辑推理。我们提出了Shortcut-Aware Reasoning Training (SART),一种梯度感知框架,通过ShortcutScore和梯度手术检测并缓解促进捷径的样本。我们的方法通过梯度与验证目标的不一致性和答案标记集中性来识别捷径信号,并相应修改训练动态。在受控推理基准上的实验表明,SART在最强基线基础上实现了+16.5%的准确率和+40.2%的鲁棒性,显著提高了在分布偏移下的泛化能力。代码可在:https://github.com/fuyanjie/short-cut-aware-data-centric-reasoning 获取。

英文摘要:

Large language models exhibit strong reasoning capabilities, yet often rely on shortcuts such as surface pattern matching and answer memorization rather than genuine logical inference. We propose Shortcut-Aware Reasoning Training (SART), a gradient-aware framework that detects and mitigates shortcut-promoting samples via ShortcutScore and gradient surgery. Our method identifies shortcut signals through gradient misalignment with validation objectives and answer-token concentration, and modifies training dynamics accordingly. Experiments on controlled reasoning benchmarks show that SART achieves +16.5% accuracy and +40.2% robustness over the strongest baseline, significantly improving generalization under distribution shifts. Code is available at: https://github.com/fuyanjie/short-cut-aware-data-centric-reasoning.

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。

↑