Guides
The blog holds my personal essays and research notes. This page holds something different: multi-part illustrated guides where each part builds on the last, every figure comes from a real seeded experiment, and many of the figures run live in your browser. Start at part one, or jump to the part you need.
Which training examples taught a model what it knows? This nine-part guide works up from the exact but unaffordable answer (leave-one-out retraining) to the estimators that scale to frontier models, with interactive figures backed by real experiments at every step. It is in beta and still actively evolving, so expect the occasional rough edge between visits.
Which training example taught the model that? The attribution question and its exact answer: leave-one-out retraining, live in your browser, and the n-squared bill that makes it unaffordable.
The influence function How one trained model answers all sixty counterfactuals at once: a 1974 idea from robust statistics, one Newton step, and the leverage it quietly drops.
When influence breaks Deep networks break the assumptions under the influence function. An autopsy in four causes, and the reframe that rescues the formula by changing the question.
Backprop through training Training is just a big computation, so differentiate it: unrolling, TracIn, SOURCE, and MAGIC as one formula with different pieces crossed out.
Influence functions at LLM scale The Hessian of a frontier model has more entries than any machine has bytes. Three ways to survive it: factor it, never form it, or let training do the damping.
Datamodels, TRAK, and the attribution landscape The map of the field: three kinds of attribution, Shapley values as linear datamodels, TRAK’s three-step recipe, and every method priced on one axis.
What is data attribution actually for? Debugging, dataset selection, poisoning, and unlearning are one trick used four ways: rewrite the task in terms of model outputs, then substitute the attribution estimate.
Shapley without the combinatorics Monte Carlo makes Data Shapley measurable, then a single-run Taylor estimate reaches nearly the same ranking for a fraction of the cost.
No method dominates: the bake-off Seven attribution scorers face the same retraining test. The rankings disagree, the baselines compete, and no method wins everywhere.
Related research: the guide’s methods are the machinery behind my COLM 2026 paper, Capability Provenance in Language Models.