Title: Mid-Think: Training-Free Intermediate-Budget Reasoning via Token-Level Triggers

Abstract: While hybrid reasoning language models are typically governed by high-level "Think" or "No-think" directives to manage their cognitive processes, our investigation reveals that these mode switches are primarily propelled by specific trigger tokens rather than the instructions themselves. Utilizing attention analysis and controlled prompting experiments, we identify that a leading "Okay" token activates reasoning capabilities, whereas the newline sequence following "" inhibits it. Leveraging this insight, we introduce Mid-Think, a straightforward, training-free prompting framework that integrates these triggers to facilitate intermediate-budget reasoning. This approach consistently surpasses fixed-token and prompt-based baselines in balancing accuracy and length. Additionally, when applied to Reinforcement Learning (RL) training following Supervised Fine-Tuning (SFT), Mid-Think cuts training duration by roughly 15% and boosts the final performance of Qwen3-8B on the AIME benchmark from 69.8% to 72.4%, and on GPQA from 58.5% to 61.1%. These results underscore the method's efficacy in both inference-time control and RL-driven reasoning training.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC