This post presents some motivation on why we work on model diffing, some of our first results using sparse dictionary methods and our next steps. This work was done as part of the MATS 7 extension. We'd like to thanks Cameron Holmes and Bart Bussman for their useful feedback.

Could OpenAI have avoided releasing an absurdly sycophantic model update? Could mechanistic interpretability have caught it? Maybe with model diffing!

Model diffing is the study of mechanistic changes introduced during fine-tuning - essentially, understanding what makes a fine-tuned model different from its base model internally. Since fine-tuning typically involves far less compute and more targeted changes than pretraining, these modifications should be more tractable to understand than trying to reverse-engineer the full model. At the same time, many concerning behaviors (reward hacking, sycophancy, deceptive alignment) emerge during fine-tuning[1], making model diffing potentially valuable for catching problems before deployment. We investigate [...]

---

Outline:

(01:39) TL;DR

(03:02) Background

(03:58) The Problem: Most Model-Specific Latents Aren't

(05:45) The Fix: BatchTopK Crosscoders

(07:17) Maybe We Don't Need Crosscoders?

(09:31) diff-SAEs Excel at Capturing Behavioral Differences

(13:00) Conclusion

(15:24) Appendix

(15:37) A) Other techniques in the model diffing toolkit

(18:55) B) Closed Form Solution for Latent Scaling

The original text contained 10 footnotes which were omitted from this narration.

---

First published:
June 30th, 2025

Source:
https://www.lesswrong.com/posts/xmpauEXEerzYcJKNm/what-we-learned-trying-to-diff-base-and-chat-models-and-why

---

Narrated by TYPE III AUDIO.

---

Images from the article:

$Frequency distribution of BatchTopK crosscoder latents. “Chat-only” latent are those with <span>__T3A_INLINE_LATEX_PLACEHOLDER___\nu\leqslant0.3___T3A_INLINE_LATEX_END_PLACEHOLDER__</span>.$

Latent frequency distributions for chat-SAE (left) and diff-SAE (right). SAE

__T3A_INLINE_LATEX_PLACEHOLDER___\nu \leqslant 0.3___T3A_INLINE_LATEX_END_PLACEHOLDER__) show similar frequency distributions to shared latents, unlike in crosscoders." style="max-width: 100%;" />

$Number of latents with <span>__T3A_INLINE_LATEX_PLACEHOLDER___\nu\leqslant x___T3A_INLINE_LATEX_END_PLACEHOLDER__</span> across dictionaries.$

Technical diagram showing neural network architecture with model layers and activations. The diagram illustrates parallel paths for chat and base models, with activations merging through an approximation layer and KL divergence calculation.

Two bar graphs comparing recovery metrics across three different methods: BatchTopK Crosscoder, SAE, and SAE on the diff. The left graph shows

Technical diagram showing AI language model latent spaces and steering capabilities.

The image displays multiple code-like text boxes demonstrating how different

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

“What We Learned Trying to Diff Base and Chat Models (And Why It Matters)” by Clément Dumas, Julian Minder, Neel Nanda

LessWrong (30+ Karma)