From Token to Action: State Machine Reasoning to Mitigate Overthinking in Information Retrieval

2025-05-29

AI Summary

In information retrieval, large language models (LLMs) employing chain-of-thought (CoT) prompting often suffer from “over-reasoning”—generating excessively long, semantically redundant reasoning paths that incur substantial computational overhead with minimal gains. To address this, we propose a state-machine-based discrete-action reasoning framework—Refine/Rerank/Stop—that explicitly models reasoning as a controlled state-transition process, enabling fine-grained trajectory control and dynamic early stopping for the first time. Our approach requires no task-specific fine-tuning and achieves zero-shot generalization across diverse LLMs and retrievers. Evaluated on the BEIR and BRIGHT benchmarks, it improves nDCG@10 by 3.4% while reducing token consumption by 74.4%, effectively mitigating both redundant reasoning trajectories and intent drift—the two core challenges in LLM-based retrieval.

Technology Category

• Knowledge Representation and Reasoning: Computational Complexity of Reasoning
• Machine Learning: Large Multimodal Models (LMMs)
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Chain-of-Thought (CoT) prompting enables complex reasoning in large language models (LLMs), including applications in information retrieval (IR). However, it often leads to overthinking, where models produce excessively long and semantically redundant traces with little or no benefit. We identify two key challenges in IR: redundant trajectories that revisit similar states and misguided reasoning that diverges from user intent. To address these, we propose State Machine Reasoning (SMR), a transition-based reasoning framework composed of discrete actions (Refine, Rerank, Stop) that support early stopping and fine-grained control. Experiments on the BEIR and BRIGHT benchmarks show that SMR improves retrieval performance (nDCG@10) by 3.4% while reducing token usage by 74.4%. It generalizes across LLMs and retrievers without requiring task-specific tuning, offering a practical alternative to conventional CoT reasoning. The code and details are available at https://github.com/ldilab/SMR.

Problem

• Mitigates overthinking in LLM reasoning for IR
• Reduces redundant and misguided retrieval trajectories
• Improves retrieval efficiency with lower token usage

Innovation

• State Machine Reasoning for efficient retrieval
• Discrete actions: Refine, Rerank, Stop
• Reduces token usage by 74.4%