Thinkfuse improves small reasoning models by selective fusion intervention
ThinkFuse: Trajectory-Aware Test-Time Fusion for Small Reasoning Models
Artificial IntelligenceComputation and Language
Summary
Small reasoning models can solve hard problems by thinking through many steps, but they sometimes get stuck on wrong ideas. The authors propose ThinkFuse, a way to help these models fix their mistakes by carefully checking if their reasoning is unstable and mixing in better ideas only when needed. This method makes the models more accurate and efficient without extra training. ThinkFuse was tested on math and knowledge tests and worked well with different models.
What this means in practice
- •For ai system developers: Improve accuracy and efficiency of reasoning in AI models by selectively fusing alternative reasoning paths during inference.
- •For intelligent tutoring system builders: Enhance feedback and reasoning quality by integrating ThinkFuse to detect and correct unstable student model reasoning trajectories.
Authors
Myunghoon Kang, Jungseob Lee, Jaehyung Seo, Heuiseok Lim
Abstract
Small reasoning models (SRMs) have shown strong performance on complex reasoning tasks by generating extended chain-of-thought trajectories, but they often fail to recover once their reasoning enters an erroneous path. Existing test-time fusion methods rely on local fusion signals to determine when to trigger fusion, which can be misled by transient uncertainty fluctuations and may reinforce unstable reasoning trajectories. We propose ThinkFuse, a training-free test-time fusion framework that selectively intervenes in unreliable reasoning segments. ThinkFuse compares segment-level uncertainty shifts with trajectory-level uncertainty trends to identify unstable reasoning points and fuse auxiliary reasoning paths into the primary model's trajectory. Extensive experiments demonstrate that ThinkFuse outperforms baselines on mathematical and knowledge-intensive reasoning benchmarks, with consistent gains across model-family combinations, and remains robust with a smaller primary model. Our analysis shows that ThinkFuse requires fewer fusion triggers and generates fewer tokens, highlighting the efficiency of selective triggering. Our code is available at https://github.com/js-lee-AI/ThinkFuse.