📄 中文摘要
适应性思维过程基于纠正反馈是人类学习中一项重要能力，尤其是在协作环境中。当前的大型语言模型训练模式主要依赖于对庞大、静态语料库的建模，虽然这种方法在知识获取方面有效，但忽视了模型动态适应其上下文所需的互动反馈循环。本研究提出了一种框架，将这种互动上下文学习能力视为一种独特的、可训练的技能，而非自发出现的特性。引入了一种可扩展的方法，将单轮可验证任务转化为由信息不对称驱动的多轮教学互动。研究表明，当前的旗舰模型在这方面的表现仍有提升空间。
改进来自自然语言反馈的互动上下文学习
出处: Improving Interactive In-Context Learning from Natural Language Feedback
发布: 2026年2月19日
📄 English Summary
Improving Interactive In-Context Learning from Natural Language Feedback
The ability to adapt one's thought process based on corrective feedback is crucial in human learning, particularly in collaborative environments. In contrast, the prevailing training paradigm for large language models heavily relies on modeling extensive, static corpora. While effective for knowledge acquisition, this approach neglects the interactive feedback loops necessary for models to dynamically adapt to their context. This research proposes a framework that treats interactive in-context learning as a distinct, trainable skill rather than an emergent property. A scalable method is introduced that transforms single-turn verifiable tasks into multi-turn didactic interactions driven by information asymmetry. The study demonstrates that current flagship models still have room for improvement in this area.