通过任务级自回归推理搭建know-act鸿沟
人工智能
2026-03-25 v1
摘要
大语言模型(LLM)常对输入存在缺陷或不完整时生成看似有效的答案。这并非由于知识缺失：在判别性提示下，相同的模型可大多数识别此类问题，但却未在标准生成性响应中体现这一点。这揭示了判别性识别与生成行为之间的根本性know-act鸿沟。先前工作主要在狭窄环境中表征此问题，如数学应用题或问答任务，聚焦如何整合这两种模式。本文提出了全面的分析，使用FaultyScience——一个新构建的大规模、跨学科的有缺陷科学问题基准测试。我们显示，这一鸿沟普遍存在，源于token级自回归，这种机制将任务选择(验证 vs. 回答)与内容生成 entangled，阻止了判别性知识的利用。为解决此问题，我们提出DeIllusionLLM，一种任务级自回归框架，显式建模此决策。通过自蒸馏，模型在单一主干网络中统一判别性判断与生成推理。实验上，DeIllusionLLM在自然提示下显著减少了在错误情况下仍给出答案的失败情况，同时保持一般推理性能，表明自蒸馏是搭建判别性-生成know-act鸿沟的有效且可扩展的解决方案。
关键词
引用
@article{arxiv.2603.22619,
title = {Bridging the Know-Act Gap via Task-Level Autoregressive Reasoning},
author = {Jihyun Janice Ahn and Ryo Kamoi and Berk Atil and Renze Lou and WonWoo Kang and Heehyun Park and Sarkar Snigdha Sarathi Das and Zhuoyang Zou and Xiaoxin Lu and Yusen Zhang and Asfahan Shah and Ridwanul Hasan Tanvir and Lingxiao Zhao and Hongxi Huang and Vignesh Venkatesh and Dianjun Lin and Hamid Shah and Wentao Wang and Zhanpeng Song and Joshua Reed Bassin and Dax Patel and Ishan Appareddy Agrahar and Sahil Pardasani and Xin Dong and Fatemeh Rahbari and Benjamin David Rishel and Soochan Andrew Lee and Yuv Boghani and Ali B. AlNaseeb and Pranav Suby and Seokhyeon Bae and Shreya Buddharaju and Damien Kula and Soumyadeep Das and Hanyang Frank Liu and Faye Mo and Wenpeng Yin},
journal= {arXiv preprint arXiv:2603.22619},
year = {2026}
}
备注
12 pages