Year: 2026
Type: conference-paper
Abstract: Chain-of-Thought (CoT) reasoning has driven recent gains of large language models (LLMs) on reasoning-intensive tasks by externalizing intermediate steps.However, excessive or redundant reasoning -so-called overthinking -can increase inference costs and lead LLMs toward incorrect conclusions.In this paper, we present REFRAIN (REFlective-Redundancy for Adaptive INference), a training-free framework that adaptively determines when to stop reasoning to mitigate overthinking.REFRAIN integrates a two-stage stop discriminator to identify reflective yet redundant reasoning and a sliding-window Upper Confidence Bound (SW-UCB) multi-armed bandit controller to dynamically adjust stopping thresholds according to problem difficulty without supervision or fine-tuning.Across four representative benchmarks and two model families, REFRAIN reduces token usage by 20-55% while maintaining or improving accuracy compared to standard CoT prompting.Extensive ablation and robustness analyses demonstrate its stability across models, scorers, and prompt variations.In summary, our findings highlight when-tostop as a new and practical axis of test-time scaling -enabling models to reason not just more, but just enough.
Authors: Renliang Sun, Wei Cheng, Dawei Li, Haifeng Chen, Wei Wang
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Awards: III: Medium: Collaborative Research: Collaborative Machine-Learning-Centric Data Analytics at Scale, CONNECT: Collaborative Network for Nurturing Ecosystems of Common Fund Team Science, Collaborative Research: III: Medium: VirtualLab: Integrating Deep Graph Learning and Causal Inference for Multi-Agent Dynamical Systems, Building BRIDGES: Coordinating the Dissemination and Training for Biomedical Artificial Intelligence
OpenAlex ID: W4417099852