Hacker NewsMar 23, 2025

Tied Crosscoders: Tracing How Chat LLM Behavior Emerges from Base Model

https://www.lesswrong.com/posts/3T8eKyaPvDDm2wzor/research-question

#HackerNews #TiedCrosscoders #ChatLLM #BaseModel #AIResearch #MachineLearning

Tied Crosscoders: Explaining Chat Behavior from Base Model — LessWrong

Abstract We are interested in model-diffing: finding what is new in the chat model when compared to the base model. One way of doing this is training…