Train on prefill, monitor at decode: Using sparse autoencoders for topic monitoring
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
SAE-feature topic monitors trained only on prefill activations generalized to per-token monitoring during decoding on Gemma-3 1B and 4B across seven topics, with labels able to switch mid-response. Practically, this means you can train cheaper offline classifiers on prompt/prefill traces and deploy compact runtime monitors during generation; without SAE features, matching performance required classifiers several times larger, and common perturbations cost at most 5 macro-F1 points.
Topic classifiers trained on sparse autoencoder features during prefill can generalize to decoding, enabling compact topic monitors that can be trained inexpensively and deployed during generation, reducing the size of classifiers by several times. This allows for more efficient and cost-effective topic monitoring in production LLMs. It enables shipping more capable and fine-grained topic monitoring with less computational overhead.