Reproduce “Train on Prefill, Monitor at Decode: Using Sparse Autoencoders for Topic Monitoring” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Rishikesh Mallagundla, Pranav Patel
TopicScope monitors topic changes token by token with a causal detector trained on labeled conversations. It applies the detector to a generating model’s internal signals while the backbone and sparse autoencoders remain frozen. On Gemma 3 4B, label-based feature selection raises held-out macro-F1 from 87.4% to 96.4% at a fixed budget. Decode purity reaches 95%, without verifying exact topic boundaries.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.