Exploring Recursive Doubt in Large Language Models

Yibo Miao, Yichuan Cao, Xiao-Shan Gao, Yinpeng Dong

Keywords: recursive doubt, feedback-guided iterative induction, obsessive cognitive tokens

Abstract: Humans sometimes experience self-doubt, repeatedly questioning their reasoning, decisions, or memories. In obsessive-compulsive disorder (OCD), this becomes a self-reinforcing loop of doubt and compulsion that leads to decision paralysis. Motivated by this analogy, we investigate whether large language models (LLMs) can exhibit a similar phenomenon, which we term \textit{Recursive Doubt}. While prior work shows that self-reflection on chain-of-thought (CoT) can improve reasoning but sometimes causes overthinking, recursive doubt represents a more pathological form of recursive reasoning that remains unexplored. In this paper, we introduce Feedback-guided Iterative iNDuction (FIND) for inducing recursive doubt. FIND leverages an auxiliary LLM to generate an induction prefix, which is optimized by the feedback of the target LLM. To understand the phenomenon, we then identify a distinctive fence-like attention pattern in certain tokens -- Obsessive Cognitive Tokens -- that repeatedly trigger self-reflection. Based on this analysis, we propose a mitigation strategy that dynamically adjusts their attention weights to suppress recursive doubt. Extensive experiments across multiple model architectures and datasets validate the effectiveness of both our induction and mitigation approaches.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 2207