Cross-Model Activation Generalizability Isn't Strong (Yet) — LessWrong

TL;DR * Tested activation similarities across different LLM families (Llama, Gemma, Qwen, Pythia) at small scale (1~3B) * CKA Similarity : Cross-ar…

x Cross-Model Activation Generalizability Isn't Strong (Yet) — LessWrong Interpretability (ML & AI) AI Frontpage 6 Cross-Model Activation Generalizability Isn't Strong (Yet) by J Lee 6th Apr 2026 17 min read 0 6 TL;DR Tested activation similarities across different LLM families (Llama, Gemma, Qwen, Pythia) at small scale (1~3B) CKA Similarity : Cross-architectural activation similarity is statistically real, but weak. Within-family activations are much stronger (4~9x) Linear Transferability : Trained linear bridges for linear activation transfers for binary classification and next token predic

related reading

• The Big LLM Architecture Comparisonmagazine.sebastianraschka.com
• Can activation verbalizers surface an internal chain of thought? — LessWronglesswrong.com
• How well do truth probes generalise? — LessWronglesswrong.com
• Large Language Diffusion Modelsarxiv.org
• Cross-Model KV Cache Transfer in LLM Families:A Closed-Form Linear Mapping for Prefill Reusearxiv.org
• Scaling Activation Oracles to Trillion-Parameter Modelstransluce.org
• [2608.21664] Measuring Activation Control in Large Language Modelsarxiv.org
• Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations — LessWronglesswrong.com
• [2512.15674] Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainersarxiv.org
• Dreaming Vectors: Gradient-descented steering vectors from Activation Oracles and using them to Red-Team AOs — LessWronglesswrong.com
• Kimi Linear: An Expressive, Efficient Attention Architecturearxiv.org
• [2607.14111] Introspection Fine-Tuning (IFT): Training Small LLMs to Introspectarxiv.org