Mid-Think

Training-free intermediate-budget reasoning through token-level triggers that induce or suppress thinking.

From Two Modes to a Useful Middle Ground

Reasoning models are usually presented with a binary choice: use full think mode or suppress reasoning with no-think. This is a coarse interface. Full reasoning can be expensive, while direct answering can leave accuracy on the table. Mid-Think: Training-Free Intermediate-Budget Reasoning via Token-Level Triggers searches for an intermediate operating point without training a new model or imposing a fixed token cutoff.

Mid-Think overview and accuracy-length trade-off
Mid-Think combines reasoning-activating and reasoning-suppressing cues to reach an intermediate accuracy–length trade-off.

Token-Level Triggers

Attention analysis reveals that high-level instructions are not the whole story. A small number of opening tokens behave like switches: a leading “Okay” token strongly attracts attention and activates reasoning, while the newline pattern after </think> suppresses it. Controlled prompts reproduce the same behavior across reasoning settings, suggesting that the model has learned lexical triggers from highly regular training templates.

Attention heat maps showing reasoning trigger tokens
Across five opening formats, generation attention concentrates on “Okay” in reasoning modes and on the newline after the closing think tag in no-think mode.

Mid-Think combines both cues in one format. The suppressing pattern first closes an empty thinking block; a new reasoning block then begins with the activating cue. This small prompt-level intervention produces an intermediate reasoning budget dynamically, rather than forcing every problem into the same preselected token limit.

Accuracy, Efficiency, and RL Training

On MATH500, Mid-Think reaches 92.1% accuracy with 2,589 average tokens, compared with 86.3% and 899 tokens for no-think, and 94.4% and 4,904 tokens for full think. Across MATH500, AIME, and GPQA, it consistently improves the accuracy–length trade-off over fixed-token and instruction-based budget controls.

Mid-Think accuracy and length trade-off against fixed reasoning budgets
Unlike a fixed token cap, Mid-Think adapts its realized budget across examples and remains close to the accuracy–length Pareto frontier.

The same mechanism also helps reinforcement-learning post-training. Applying Mid-Think after supervised fine-tuning reduces RL training time by about 15% (54 to 46 hours for Qwen3-8B), while improving final think-mode accuracy from 69.8% to 72.4% on AIME and from 58.5% to 61.1% on GPQA.

Entropy by reinforcement-learning training step
Entropy over GRPO training steps.
Entropy by reinforcement-learning wall-clock time
Mid-Think reaches the end of training in less wall-clock time.

The result is a lightweight way to regulate reasoning at both inference and training time. The paper was published in Findings of ACL 2026.