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

TL;DR

The paper identifies that mode-switching in hybrid reasoning LLMs is driven by token-level triggers rather than high-level prompts. It introduces Mid-Think, a training-free prompting format that blends a reasoning-activating cue ('Okay') with a suppression cue (the '</think>' newline) and an explicit <reason> delimiter to realize intermediate-budget reasoning. Mid-Think achieves Pareto-optimal accuracy-efficiency trade-offs across benchmarks such as MATH500, AIME, and GPQA and yields substantial RL-training efficiency gains after SFT, e.g., improving AIME accuracy from $69.8\%$ to $72.4\%$ and GPQA from $58.5\%$ to $61.1\%$, while reducing training time by about $15\%$. This approach provides practical gains for both inference-time control and RL-based reasoning training, though it relies on preexisting token-level cues and offers limited dynamic budget control.

Abstract

Hybrid reasoning language models are commonly controlled through high-level Think/No-think instructions to regulate reasoning behavior, yet we found that such mode switching is largely driven by a small set of trigger tokens rather than the instructions themselves. Through attention analysis and controlled prompting experiments, we show that a leading ``Okay'' token induces reasoning behavior, while the newline pattern following ``</think>'' suppresses it. Based on this observation, we propose Mid-Think, a simple training-free prompting format that combines these triggers to achieve intermediate-budget reasoning, consistently outperforming fixed-token and prompt-based baselines in terms of the accuracy-length trade-off. Furthermore, applying Mid-Think to RL training after SFT reduces training time by approximately 15% while improving final performance of Qwen3-8B on AIME from 69.8% to 72.4% and on GPQA from 58.5% to 61.1%, demonstrating its effectiveness for both inference-time control and RL-based reasoning training.

Figures (9)

• Figure 1: Illustration of Mid-Think and its performance on MATH500. (a) Overview of Mid-Think comparing with Think and No-Think. In the Think mode, subsequent tokens primarily attend to the cue token "Okay", while in the No-think mode, generated tokens focus on the newline following the </think> marker (i.e., the \\ n\\ n token). Mid-Think combines both cues into a unified prompting format to induce intermediate reasoning behavior. (b) Accuracy and (c) average output length under No-think, Think, and Mid-Think settings. Mid-Think achieves intermediate-budget reasoning without additional training, retaining most accuracy gains while significantly reducing generation length.
• Figure 2: Average attention from generated tokens to different opening tokens under five generation modes. Darker red indicates higher attention received from subsequent tokens. When “Okay” appears at the beginning, the model produces an explicit reasoning process ("wait", "alternatively", etc), and attention is primarily concentrated on “Okay”. In the No-Think mode, the \\ n\\ n following </think> absorbs most of the attention mass.
• Figure 3: Budget-controlled reasoning on MATH500 using Qwen3-14B. The figure reports the average length and accuracy under different reasoning budgets. Both metrics increase steadily as the budget grows, validating the effectiveness of the budget-control mechanism.
• Figure 4: Overview of the budget-controlled method. The model first generates a full response. The reasoning (think) content is then truncated to the specified budget (in tokens) and concatenated with the remaining prompt, after which the model generates the final response.
• Figure 5: Comparison between hybrid-thinking models (from Qwen3 family) under different reasoning budgets and the proposed Mid-Think across multiple datasets. Results are reported for Qwen3-8B and Qwen3-14B. We evaluate these models at varying budgets and Mid-Think with different tags (<reason>, <begin>, <less think>). Across datasets (MATH500, AIME and GPQA), Mid-Think consistently exhibits performance corresponding to intermediate reasoning budgets, and on GPQA it even surpasses fixed-budget baselines, achieving Pareto-optimal trade-offs between Think and No-Think mode.