📄 中文摘要
长上下文处理仍然是语言模型面临的核心挑战：即使在扩展的上下文窗口下，模型也常常无法可靠地提取、推理和使用长上下文中的信息。递归语言模型（RLM）通过在推理过程中以程序化的方式将长上下文分解为递归子调用，尝试解决这一挑战。RLM的成功在很大程度上依赖于如何选择这些上下文交互程序，而这一点尚未得到充分探索。研究提出了SRLM框架，通过不确定性感知的自反机制增强程序化上下文交互。SRLM利用了三种内在信号：自我协同、上下文不确定性和反馈机制，从而提高了长上下文处理的有效性。
递归语言模型与不确定性相遇：自反程序搜索在长上下文中的意外有效性
📄 English Summary
Recursive Language Models Meet Uncertainty: The Surprising Effectiveness of Self-Reflective Program Search for Long Context
Long-context handling remains a core challenge for language models, as they often struggle to reliably extract, reason over, and utilize information across extended contexts. Recursive Language Models (RLM) have attempted to address this challenge by decomposing long contexts into recursive sub-calls through programmatic interaction during inference. However, the success of RLM critically hinges on the selection of these context-interaction programs, a largely unexplored area. This research introduces the SRLM framework, which enhances programmatic context interaction with uncertainty-aware Self-Reflection. SRLM leverages three intrinsic signals: self-coherence, context uncertainty, and feedback mechanisms, thereby improving the effectiveness of long-context handling.