EP 02 · How AI Actually Works
Why AI Says “Wait”: How Thinking Models Actually Work (Every Step Measured)
What the Thinking toggle really does, why models say “Wait”, when more thinking hurts, and how GRPO teaches reasoning.
TL;DR
Thinking mode is the same model with a different prompt: turning it off just pre-fills 4 tokens that close the <think> box. The scratchpad helps because written words act as working memory (99.99% accuracy with real reasoning, 35.7% with filler dots), but it can overthink, costs 147,456 bytes of KV-cache per token, and is taught with reinforcement learning such as GRPO.
- 4 tokensthe whole toggle
- 99.99%real scratchpad
- 35.7%filler dots
- 147,456 BKV-cache per token
- 7.8% → 66.5%GRPO, 250 steps
Summary
Part 2 opens the <think> scratchpad of Qwen3-4B. In one captured trace, the model spelled “strawberry”, found r's at positions 3, 8 and 9, wrote that there were two, then sampled the token “ Wait” (#13824), recounted and answered 3. The article asks where that self-correction comes from, using 696+ benchmark completions and its own GRPO training run.
The first finding is that Thinking mode is not a second model. Both modes use the same 4,022,468,096 parameters; switching thinking off just pre-fills four tokens, an empty <think></think> block, so the model skips straight to the answer.
Writing is what helps. A probe of the model's confidence in “3” climbs to 99.0% once the word is spelled out, drops to 22% at the slip and rebounds to 99.8%. Replacing the thinking text with the same number of filler dots scores 35.7%, barely above 33.6% with no thinking, against 99.99% for real reasoning. Words like “Wait” are high-entropy forking tokens, and “Wait” only becomes the top candidate at Layer 33 of 36.
More thinking is not always better. On GSM8K, thinking on (93.9%) and step-by-step answers with thinking off (90.6%) are within noise, while AIME 2024 jumps from 25.0% to 73.8%. A bat-and-ball riddle falls from 100% to 37.5% with thinking on. Every hidden token also costs 147,456 bytes of KV-cache. Finally, 250 GRPO steps lifted Qwen3-0.6B-Base from 7.8% to 66.5% on held-out GSM8K, while Qwen3-4B's habit of saying “Wait” came from distillation from larger teachers.
This page is a short summary written for this site. Every figure, table, code sample and reference is in the original article on Towards AI.
Questions this article answers
What does the Thinking toggle actually do in Qwen3?
It changes the prompt, not the model. Qwen3-4B uses the same file and the same 4,022,468,096 parameters in both modes. With thinking off, the chat template appends four tokens, [151667, 271, 151668, 271], which spell an empty <think></think> block. Seeing the box already closed, the model goes straight to its final answer. With thinking on, it opens the scratchpad itself and wrote a median of 1,370 thinking tokens in the tests.
Why do reasoning models say “Wait”?
Nobody hard-codes it. In the measured traces, ordinary calculation tokens had a median entropy of just 0.012 bits, but discourse words such as “Wait”, “Hmm” and “But” spiked to between 0.34 and 2.90 bits. These high-uncertainty forking tokens are where a chain of reasoning can change direction. A logit lens showed “Wait” only becoming the model's top choice at Layer 33 of 36, so the doubt is computed late in the network.
Is chain-of-thought just extra compute time?
The article's filler-dot test says no. Across 107 real thinking traces, every thinking token was replaced with the same number of dots, giving the model identical forward passes. Accuracy was 99.99% with the real reasoning text, 54.0% with the words shuffled, 35.7% with dots and 33.6% with no thinking at all. What is written on the scratchpad, not the extra passes, is what does the work.
Can thinking mode make an LLM less accurate?
Yes, on easy or deceptive prompts. With thinking on, Qwen3-4B wrote 226 tokens before answering 2 + 3. On the bat-and-ball riddle it was right 64 out of 64 times with thinking off, but only 37.5% of the time with it on, after second-guessing the wording. When prompts claimed a Stanford professor had given a wrong answer, 70 of 80 runs looped until they hit the 6,144-token limit.
How much GPU memory do hidden thinking tokens use?
The same as visible ones. In Qwen3-4B, each token stores keys and values for 36 layers, 8 key-value heads and a head size of 128 in bf16, which works out to 147,456 bytes, about 144 KB. A thousand thinking tokens is roughly 140.6 MiB for one conversation. On 100 GSM8K problems, thinking mode produced 1.84 million tokens against 0.24 million without it, with 6.6× higher median latency.
What is GRPO and how does it teach models to reason?
Group Relative Policy Optimization, used by DeepSeek-R1, samples a group of answers (here 8) for one problem, scores each with an automatic checker, and pushes up tokens from answers that beat the group average while pushing down the rest, with no separate critic model. In the article's run, 250 GRPO steps, about 90 minutes on one GPU, took Qwen3-0.6B-Base from 7.8% to 66.5% on held-out GSM8K.