# KL divergence explained: measuring how much a model's answers have shifted

> Original title: KL：模型，你变了多少

- Source: Dongxi NLP
- Published: 2026-11-05T07:19:43.000Z
- Bad Influence: https://badinfluence.news/items/n52edkbqjmbvxrjadtnmjdc5n
- Original: https://x.com/dongxi_nlp/status/2118241835348893885

## Summary

The article explains KL divergence as a measure of how far a fine-tuned language model's output distribution has moved from its reference model, using a boiled-egg example with probabilities of 0.40 to 0.30 and 0.12 to 0.22. It walks through the three steps of ratio, log, and weighted average, reporting a KL of 0.047 nats for that shift. The author says KL is never negative and serves as the penalty that keeps reinforcement-learning training close to the reference model.

## Post

https://x.com/i/article/2118081140015656961