VeriThinker: Learning to Verify Makes Reasoning Model Efficient

TL;DR

VeriThinker addresses the inefficiency of long Chain-of-Thought reasoning in large models by introducing Supervised Verification Fine-Tuning (SVFT), which trains models to verify the correctness of CoT solutions rather than imitate short target chains. This reframing pushes the model to maximize $p(\text{acc}|h)$ rather than matching an explicit concise chain, enabling substantial CoT length reduction while preserving or improving accuracy. The method is evaluated across multiple long-CoT models and math benchmarks, with results showing notable token reductions (approximately 28-29%) and accuracy gains on challenging datasets such as AIME 2024/2025; additionally, Solution-wise Speculative Reasoning (SSR) leverages SVFT to dramatically boost throughput by using draft solutions from short-CoT LLMs. Overall, VeriThinker offers a practical path to efficient reasoning with scalable inference and extends to speculative decoding, supported by code release for reproducibility.

Abstract

Large Reasoning Models (LRMs) excel at complex tasks using Chain-of-Thought (CoT) reasoning. However, their tendency to overthinking leads to unnecessarily lengthy reasoning chains, dramatically increasing inference costs. To mitigate this issue, we introduce VeriThinker, a novel approach for CoT compression. Unlike conventional methods that fine-tune LRMs directly on the original reasoning task using synthetic concise CoT data, we innovatively fine-tune the model solely through an auxiliary verification task. By training LRMs to accurately verify the correctness of CoT solutions, the LRMs inherently become more discerning about the necessity of subsequent self-reflection steps, thereby effectively suppressing overthinking. Extensive experiments validate that VeriThinker substantially reduces reasoning chain lengths while maintaining or even slightly improving accuracy. When applied to DeepSeek-R1-Distill-Qwen-7B, our approach reduces reasoning tokens on MATH500 from 3790 to 2125 while improving accuracy by 0.8% (94.0% to 94.8%), and on AIME25, tokens decrease from 14321 to 10287 with a 2.1% accuracy gain (38.7% to 40.8%). Additionally, our experiments demonstrate that VeriThinker can also be zero-shot generalized to speculative reasoning. Code is available at https://github.com/czg1225/VeriThinker

Figures (7)

• Figure 1: The key distinction between VeriThinker and traditional SFT or RL-based long-to-short methods. We uniquely train LRMs on an auxiliary CoT verification task, achieving effective CoT compression without relying on synthetic target reasoning chains.
• Figure 2: (a)-(b): Token counts and accuracy in reasoning tasks across different training datasets. (c) Probability of self-reflection during reasoning for correct and incorrect solutions.
• Figure 3: Speculative reasoning results on three reasoning models. When using Qwen-2.5-Math-Instruct-7B as the draft model, most problems in MATH500 and GSM8K can be solved with short CoT model, while only a few require activation of the long CoT model for more complex solutions.
• Figure 4: Training data format comparison.
• Figure 5: Case study 1 on CoT Compression.