# Reproduction: On the Accuracy of Newton Step and Influence Function Data Attributions

[HF paper page](https://huggingface.co/papers/2512.12572)

## Pages

| Page |
| --- |
| [Executive summary](#/executive-summary) |
| [Claim 1: New analysis of influence functions and Newton step for convex ERM replaces global strong convexity assumptions with local conditions.](#/claim-1-new-analysis-of-influence-functions-and-newton-step-for-convex-erm-replaces-global-strong-convexity-assumptions-with-local-conditions) |
| [Claim 2: Bounds for logistic regression with Gaussian features capture the correct scaling up to polylogarithmic factors, yielding matching upper and lower bounds.](#/claim-2-bounds-for-logistic-regression-with-gaussian-features-capture-the-correct-scaling-up-to-polylogarithmic-factors-yielding-matching-upper-and-lower-bounds) |
| [Claim 3: Newton step shows markedly more accurate data attribution than influence functions, explaining open discrepancies between scalar causal scores and interpretability.](#/claim-3-newton-step-shows-markedly-more-accurate-data-attribution-than-influence-functions-explaining-open-discrepancies-between-scalar-causal-scores-and-interpretability) |
| [Conclusion](#/conclusion) |