Train on prefill, monitor at decode: Using sparse autoencoders for topic monitoring
Not specified in the provided content
Abstract
The paper introduces TopicScope, a framework for training causal topic classifiers using sparse autoencoder features to label topics in model responses during generation.
Reality Card
TopicScope enables the training of compact topic monitors that can be deployed during generation, achieving robust performance with minimal size.
Monitors trained on features from prefill hidden states generalize to decoding, maintaining performance with a maximum loss of 5 macro-F1 points under perturbations.
The classifiers reaching baseline performance without the sparse autoencoder were significantly larger, which may affect deployment efficiency.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.