Theory Thursday: LLMs use scratchbooks, too
Art by @basilonmypizza: https://lnkd.in/eF8FkWzN - https://basilhefti.ch/
Chances are, you started the year with good resolutions. I applaud you. And maybe you acquired a scratchbook to journal your journey. Let’s keep our fingers crossed.
It turns out that large language models like to doodle as well. As an end user, you often don’t see it. Internally some models have started to use special tokens or modes for explicit reasoning, with “thinking” being one of them. What are these for?
Models perform better if they are allowed to take notes while they work. Much like us, writing things down changes how they think. Ideas can be revisited. Intermediate steps become visible. Errors are easier to spot.
For a long time, models like GPT-4 optimised for fast thinking. Short chains of reasoning stayed implicit, not written out. This changed with recent work on inference-time scaling. In their 2025 paper “A Survey of Frontiers in LLM Reasoning”, Ke et al. show that models allowed to generate explicit intermediate reasoning outperform comparable models hidden reasoning.
This insight reshaped the field. DeepSeek-R1 marked a shift toward learning-to-reason systems, trained to value intermediate steps. Qwen3 extended this further, effectively turning the scratchbook into part of the model’s inference process (Qwen3 is also where people found additional special tokens, such as video_pad).
This is a transition: progress now comes less from bigger models and more from better thinking spaces. Time to dust your scratchbook off, then ;)
What other tools have quietly become part of your thinking?
—
• A Survey of Frontiers in LLM Reasoning” (Ke et al., 2025)
• This is the 40th post with illustrations - do you like them?
• In case you missed it: my Silvester post was a blast ;)
• Art: https://lnkd.in/eF8FkWzN https://basilhefti.ch/
