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

TL;DR

This paper introduces State Machine Reasoning (SMR), a new framework that reduces overthinking in language model-based information retrieval by controlling reasoning steps, leading to improved performance and efficiency.

Contribution

The paper presents SMR, a transition-based reasoning framework with discrete actions that mitigates overthinking in IR, demonstrating significant performance gains and token reduction without task-specific tuning.

Findings

SMR improves retrieval nDCG@10 by 3.4%.

SMR reduces token usage by 74.4%.

SMR generalizes across models and retrievers.

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.