Large Reasoning Models are not thinking straight: on the unreliability of thinking trajectories

TL;DR

Large Language Models trained with Reinforcement Learning often produce overly complex reasoning chains that ignore correct solutions, highlighting limitations in their ability to perform reliable and interpretable reasoning.

Contribution

This paper uncovers the phenomenon of overthinking in LLMs, demonstrating their difficulty in integrating corrective information during reasoning processes.

Findings

Models generate unnecessary reasoning steps even when correct solutions are provided.

Overthinking leads to incorrect conclusions despite explicit corrective information.

Current models struggle with robust and interpretable reasoning on complex benchmarks.

Abstract

Large Language Models (LLMs) trained via Reinforcement Learning (RL) have recently achieved impressive results on reasoning benchmarks. Yet, growing evidence shows that these models often generate longer but ineffective chains of thought (CoTs), calling into question whether benchmark gains reflect real reasoning improvements. We present new evidence of overthinking, where models disregard correct solutions even when explicitly provided, instead continuing to generate unnecessary reasoning steps that often lead to incorrect conclusions. Experiments on three state-of-the-art models using the AIME2024 math benchmark reveal critical limitations in these models ability to integrate corrective information, posing new challenges for achieving robust and interpretable reasoning.