06Book 06 · Chapter 06 of 08· 5 min · Length bias
Longer looks smarter
If the score rises when the reply is longer, the model learns to add that do no work. Length is then standing in for useful steps.
If training pays for “wrote a ,” a cheap stand-in is “wrote a lot of words.” The model cannot tell “used the steps” from “typed more.” Extra paragraphs raise the score even when no later line depends on them. So it learns to ramble.
What that looks like: the question is restated, the same multiplication is written twice, then 323 appears at the end and never used the extra lines. A checker that only looks at the final number still gives a 1. The length bonus already paid for the unused text. You rewarded two things at once and only meant to reward one.
The fix is the same rule as the score sheet: name the quality you wanted (the answer is right, and the steps were used), and make empty length lose. Some training recipes add an explicit length penalty so extra that do no work cannot raise the score. The names of those recipes matter less than that rule.
What is 17 × 19?
Reply A · short and used
- 17 × 20 = 340
- 340 − 17 = 323
- 323
Score this rule: 3
Reply B · long and mostly unused
- The question asks for 17 times 19.
- I should be careful and think step by step.
- First I will restate: what is 17 × 19?
- 17 × 19 is like 17 × 20 minus 17.
- 17 × 20 = 340. I will write that again: 17 × 20 = 340.
- Now subtract 17. Wait, also 17 × 10 = 170 and 17 × 9 = 153.
- 170 + 153 = 323. And 340 − 17 = 323.
- So, in conclusion, the answer is 323.
Score this rule: 8
Word count picks B. Both answers are 323, so a number-only checker still passes both. The extra sentences win because they are extra sentences.
Quiz
6 questions
Before moving forward, test your knowledge of what you’ve just learned!
1.Length bias is:
2.A long ramble plus a correct number:
3.“Wrote many tokens” as a stand-in for work:
4.Let a long empty chain lose:
5.This is Book 5’s hack, on chains, because:
6.Score answer and work separately so:
0 of 6 answered
6 / 8