Reproduce “Interpretable sparse fine-tuning with TopKLoRA enables faithful circuit discovery” 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
Marek Masiak, Andrzej Szablewski, Nicola Cancedda, Constantin Venhoff
TopKLoRA makes fine-tuning updates sparse and uses CLCD to identify adapter latents carrying a trigger-conditioned response. In models up to 7B, circuits were smaller than rank-matched dense-LoRA circuits; for Gemma-2-2B across layers 15–23, median sizes were 7.4% versus 69.4%. Ablation largely stopped the response while multi-layer models retained instruction following. The tests used a known tag trigger.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.