Part 04 · LLM Internals
Test-Time Compute and Thinking Tokens
Thinking tokensTest-time computeRL with verifiable rewardsWhen to use reasoning
Some models think before they answer: they write a long hidden scratchpad first. How that works, how reinforcement learning on checkable answers teaches it, and when the extra tokens are worth the wait and the bill.
⏱ 1-2 hours · Station 30 of 92
✎
This chapter is being written.
The route is mapped; the full walkthrough arrives here soon.