ThinkFuse is a training‑free test‑time fusion framework that intervenes only when a small reasoning model (SRM) enters an unreliable reasoning segment. It identifies unstable points by comparing segment‑level uncertainty shifts with the overall trajectory‑level uncertainty trend, and then fuses auxiliary reasoning paths into the primary model's trajectory. Compared with conventional methods that trigger fusion solely on local uncertainty signals, ThinkFuse avoids false triggers caused by transient fluctuations, thereby preserving reasoning stability.
Extensive experiments on mathematical and knowledge‑intensive reasoning benchmarks, covering various model families, show that ThinkFuse consistently outperforms baselines in accuracy and remains robust even when the primary model is reduced in size. Further analysis reveals that ThinkFuse requires fewer fusion triggers and generates fewer additional tokens, highlighting the efficiency of its selective triggering.
The code is publicly available at https://github.com/js-lee-AI/ThinkFuse
Review