Learning a Generative Meta-Model of LLM Activations
This work addresses interpretability in large language models by offering a scalable method without restrictive structural assumptions, though it is incremental as it builds on existing generative modeling approaches.
The authors tackled the problem of analyzing neural network activations by training diffusion models on one billion residual stream activations to create generative meta-models, finding that diffusion loss decreases with compute and improves fluency in steering interventions, with sparse probing scores scaling as loss decreases.
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.