Learning a Generative Meta-Model of LLM Activations

TL;DR

This paper introduces generative meta-models trained on neural network activations, which improve interpretability and intervention quality without relying on structural assumptions, by modeling internal states with diffusion models.

Contribution

It demonstrates that diffusion-based meta-models trained on large-scale activations can enhance interpretability and intervention fidelity in neural networks.

Findings

Diffusion loss decreases smoothly with compute resources.

Meta-models improve intervention fluency as loss decreases.

Neurons increasingly isolate concepts into individual units.

Abstract

Existing approaches for analyzing neural network activations, such as PCA and sparse autoencoders, rely on strong structural assumptions. Generative models offer an alternative: they can uncover structure without such assumptions and act as priors that improve intervention fidelity. We explore this direction by training diffusion models on one billion residual stream activations, creating "meta-models" that learn the distribution of a network's internal states. We find that diffusion loss decreases smoothly with compute and reliably predicts downstream utility. In particular, applying the meta-model's learned prior to steering interventions improves fluency, with larger gains as loss decreases. Moreover, the meta-model's neurons increasingly isolate concepts into individual units, with sparse probing scores that scale as loss decreases. These results suggest generative meta-models offer a scalable path toward interpretability without restrictive structural assumptions. Project page: https://generative-latent-prior.github.io.