Reproduce “Data Attribution via Sketched Metadifferentiation” 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
Yuxi Chen, Hamza Golubovic, Han Tong, Arian Maleki, Andrew Ilyas
The authors introduce MAGE and SPELL to estimate how training examples affect many held-out queries using limited training replays. MAGE targets influence-matrix reconstruction, while SPELL targets equal-query counterfactual attribution; each replay is paired with a forward-mode pass. On CIFAR-10, MAGE’s error was 0.4872 versus PCA’s 0.5316, while SPELL’s LDS was 0.3353 versus 0.2461 for random probes at 1% removal; these results do not establish wall-clock speedups or universal superiority.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.