Let LRMs Break Free from Overthinking via Self-Braking Tuning

TL;DR

The paper tackles the cost of overthinking in large reasoning models by introducing Self-Braking Tuning (SBT), an endogenous framework that teaches models to self-regulate reasoning length. It defines two quantitative signals—the reasoning efficiency ratio $\eta_s$ and the overthinking marker ratio $\kappa_t$—and combines them into an Overthink Score with $\beta=0.1$ to identify redundant thinking. Two data-construction strategies, SBT-E (exact) and SBT-D (dynamic), along with a braking-prompt mechanism and a self-regulating training regime (masked redundant thinking plus natural-language braking cues), are used to train models on $OpenR1$-Math trajectories. Empirical results on GSM8K, MATH500, AMC23, and AIME show token reductions of roughly $30$–$60\%$ with accuracy largely preserved, with larger gains for general-purpose models and strong cross-domain transfer to non-mathematical benchmarks. Overall, SBT demonstrates that LRMs can autonomously recognize when further reasoning is unproductive, enabling more efficient and scalable reasoning without external controls.

Abstract

Large reasoning models (LRMs), such as OpenAI o1 and DeepSeek-R1, have significantly enhanced their reasoning capabilities by generating longer chains of thought, demonstrating outstanding performance across a variety of tasks. However, this performance gain comes at the cost of a substantial increase in redundant reasoning during the generation process, leading to high computational overhead and exacerbating the issue of overthinking. Although numerous existing approaches aim to address the problem of overthinking, they often rely on external interventions. In this paper, we propose a novel framework, Self-Braking Tuning (SBT), which tackles overthinking from the perspective of allowing the model to regulate its own reasoning process, thus eliminating the reliance on external control mechanisms. We construct a set of overthinking identification metrics based on standard answers and design a systematic method to detect redundant reasoning. This method accurately identifies unnecessary steps within the reasoning trajectory and generates training signals for learning self-regulation behaviors. Building on this foundation, we develop a complete strategy for constructing data with adaptive reasoning lengths and introduce an innovative braking prompt mechanism that enables the model to naturally learn when to terminate reasoning at an appropriate point. Experiments across mathematical benchmarks (AIME, AMC, MATH500, GSM8K) demonstrate that our method reduces token consumption by up to 60% while maintaining comparable accuracy to unconstrained models.

Figures (3)

• Figure 1: Demonstration of Self-Braking Tuning Effectiveness. In the single-example case (a), the self-braking tuned model exhibits spontaneous termination of overthinking and significantly reduces token usage. On major mathematical benchmarks (b), compared to using OpenR1-Math openr1math220k as the SFT dataset, the self-braking tuned Qwen2.5-Math-1.5B-Instruct yang2024qwen2 achieves a substantial reduction in tokens consumed during inference while maintaining comparable accuracy.
• Figure 2: Overview of Self-Braking Tuning. Left: Data construction process with overthinking identification and self braking truncation strategies. Right: An example of automatic reasoning termination in a trained Self-Braking LLM.
• Figure 3: In panel (a), we present a representative example from DeepSeek-R1-Distill-Qwen-7B. In panel (b), we analyze the model’s performance across GSM8K, MATH500, and AIME benchmarks, showing that the Foundation Solution plays a critical role across tasks of varying difficulty.