Don’t Overthink It: Detecting and Truncating Overthinking in LRMs with Lightweight DeBERTa

Hsien Xin Peng, Aaroosh Rustagi, Khushal Murthy, Attrey Koul, Kevin Zhu

Keywords: Language Reasoning Model, Deepseek, DeBERTa, Reasoning, CoT

Abstract: Recently, Language Reasoning Models (LRMs) have seen huge improvements in their problem-solving skills due to their ability to effectively utilize Chain-of-Thought (CoT) to explain their reasoning before committing to a final answer. However, these explanations often become unnecessarily verbose and redundant, leading to significant computational overhead. We propose a framework that segments model output into distinct reasoning steps, and train a lightweight encoder-only model to predict whether each reasoning step is useful to solving the problem or not. Moreover, we show that our lightweight model generalizes across held-out datasets and models without retraining or finetuning, allowing for seamless integration with existing LRMs. Experiments show that by varying the number of consecutive non-useful steps allowed before a forced early-exit, our framework provides a substantial reduction in tokens by 23.3\%.

Submission Number: 31