Large Reasoning Models are not thinking straight: on the unreliability of thinking trajectories
This work highlights critical limitations in achieving robust and interpretable reasoning for AI systems, which is an incremental but important step for improving reliability in real-world applications.
The paper investigates the problem of overthinking in large language models, where they generate unnecessary reasoning steps that lead to incorrect conclusions, even when correct solutions are provided, as shown by experiments on the AIME2024 math benchmark with three state-of-the-art models.
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.