07Book 06 · Chapter 07 of 08· 5 min · Steps that do not do anything
Theater versus work
Unused steps look like reasoning and leave the final answer the same if you delete them. Used steps are lines the later text depends on: take one out, and the final answer changes.
The model can copy the look of reasoning because that look won votes. “Let’s think step by step,” then a number that was going to be guessed anyway. A checker that only reads the final number still gives a 1. The middle lines never fed the answer. That is unused text wearing a reasoning opening.
Used work has a simple test. Later lines actually need the earlier ones — the way “340 − 17 = 323” needs “17 × 20 = 340” to already be on the page. Delete a used line and the next-word list for the answer changes, so the landing often changes. Delete an unused line and the landing stays. The unused line was never in the history that mattered.
If the score only asks “did this reply include steps?” the model learns to include steps. It does not learn to use them. You get openings and recaps that look careful and do not change the number. Score both: the landing, and whether the lines before it were actually used.
Quiz
6 questions
Before moving forward, test your knowledge of what you’ve just learned!
1.Theater is:
2.Work is:
3.If you could delete a line and the answer would match:
4.Scoring only “has a chain”:
5.Voters liking the costume:
6.Attention is how you would see use, because:
0 of 6 answered
7 / 8