How do reasoning models "think"? What are thinking tokens?
LLM & AILevel: FoundationWhat Are Reasoning Models
How do reasoning models "think"? What are thinking tokens?
Hints
- Identify the stage: pretraining, adaptation, generation or evaluation.
- Name the inputs, constraints and a case where the approach fails.
Answer
Thinking tokens are tokens generated by the model during a private reasoning phase before producing the final response. In most implementations:
- Model receives user query
- Model generates a
<thinking>block: scratch-pad reasoning, hypothesis testing, self-correction - Model generates the final
<answer>based on its thinking
User: "If a train leaves Chicago at 9am going 60mph and another leaves NY at 10am going 80mph..."
<thinking>
Let me set up equations. Distance between Chicago and NY is ~790 miles.
Train 1: starts at t=0, position = 60t
Train 2: starts at t=1, position = 790 - 80(t-1)
They meet when: 60t = 790 - 80(t-1)
60t = 790 - 80t + 80
140t = 870
t = 6.21 hours after 9am = 3:13pm
Let me verify: Train 1 at 6.21h = 372.9 miles, Train 2 traveled 5.21h × 80 = 416.7...
hmm 372.9 + 416.7 = 789.6 ≈ 790 ✓
</thinking>
The trains meet at approximately 3:13 PM Central Time.
The thinking tokens are often not shown to users (hidden) — they're internal scratch space. This has implications for API costs (you may be billed for invisible tokens).
Source and licence: JustInCache & contributors — Awesome AI Interviews · Apache-2.0