Reproduce “Hippocampal consolidation without replay in fully fine-tuned networks” 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

Keuntae Kim, Yong Suk Choi

DeltaHippo addresses class-incremental learning by steering later weight updates with a compact training-only index of earlier network activity. It adds class-conflict responses when new and earlier classes overlap; the index is removed for inference. In Llama-3.2-1B on TACRED, absorption among eight earlier near-duplicate classes was 2.8%, versus 9.0% for frozen-backbone SEQ*. Near-duplicates followed TACRED label-string rules.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.