Learning a Generative Meta-Model of LLM Activations
Source: https://x.com (tweet) Author: Stanford/Berkeley team (Alec Radford involved) Date: 2025-2026
Summary
Diffusion models trained on a billion LLM activations to learn a generative model of the activation space. Enables sampling new activation patterns and studying the structure of LLM internals. Joint work including Alec Radford, Trevor Darrell, Jacob Steinhardt.
Key Claims
- Trained diffusion models on 1 billion LLM activations
- Result: generative meta-model of the activation distribution
- Enables: sampling new activations, studying structure of LLM internals
- Uses: interpretability, studying model behavior, controllability
- Strong team: Alec Radford, Trevor Darrell, Jacob Steinhardt
Concepts
- Mechanistic Interpretability — generative model of internal representations