BROWSE
Detail
Conference Paper
Breaking Cognitive Fixation in Multi-Turn Dialogue with Self-Distancing and Incubation
Cited - time in 
Share 
Share
- Authors
- Hyeok-Min Gwon, Jin-Xia Huang, Yohan Lee
- Issue Date
- 2026-05
- Citation
- International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2026, pp.1-5
- Publisher
- IEEE
- Language
- English
- Type
- Conference Paper
- Abstract
- Large language models (LLMs) have shown impressive capabilities across diverse natural language processing tasks. Despite their success in single-turn interactions, LLMs often struggle to maintain task success rate in multi-turn dialogue, which constitutes the core of real-world user interaction. We interpret the degradation as being caused by a cognitive bias akin to those observed in human reasoning, such as cognitive fixation, where early missteps constrain effective multi-turn reasoning. Inspired by this parallel, we incorporate insights from cognitive psychology to help models recover from earlier mistakes and engage in more flexible reasoning. To this end, we introduce two strategies: Self-Distancing, which guides the model to reassess its own outputs from an external perspective and the Incubation, which employs abstract summarization to break the fixation loop. Even with their simplicity and model-agnostic nature, these strategies yield a 8.0% - 29.0% performance recovery on multi-turn tasks compared to baseline LLMs.
- Keyword
- Large language models, Dialogue systems, Error accumulation, Error recovery, Cognitive psychology
- KSP Keywords
- Cognitive Bias, Cognitive Psychology, Error Recovery, Error accumulation, Natural Language Processing(NLP), Performance recovery, Real-world, User interaction, dialogue system, language models, task success rate