snaykey/repro-newton-step-influence-function-data-attributions

1{
2  "schema_version": 1,
3  "title": "On the Accuracy of Newton Step and Influence Function Data Attributions",
4  "emoji": "📊",
5  "space_id": "snaykey/repro-newton-step-influence-function-data-attributions",
6  "paper": {
7    "arxiv_id": "2512.12572",
8    "openreview_id": "mDo8XNqopd"
9  },
10  "tags": [
11    "icml2026-repro",
12    "paper-mDo8XNqopd"
13  ],
14  "updated_at": "2026-07-24T19:41:56+00:00",
15  "root": {
16    "slug": "index",
17    "title": "On the Accuracy of Newton Step and Influence Function Data Attributions",
18    "file": "pages/index.md",
19    "children": [
20      {
21        "slug": "claim-1-theorem-1-5-local-strong-convexity",
22        "title": "Theorem 1.5 bounds the Newton-step approximation error to the true leave-one-out retrained parameter using only local strong convexity in a neighborhood of the Newton step, rather than the global strong convexity assumed in prior analyses (Theorem 1.5, Section 3).",
23        "file": "pages/claim-1-theorem-1-5-local-strong-convexity/page.md",
24        "children": []
25      },
26      {
27        "slug": "claim-2-theorem-1-2-error-scaling-laws",
28        "title": "For logistic regression with n samples, d-dimensional Gaussian features, and k removed points, the average-case Newton-step attribution error scales as Õ(kd/n²), while the average-case influence-function error scales as Õ((k^{3/2}d^{1/2}+k^{1/2}d^{3/2})/n²) (Theorem 1.2).",
29        "file": "pages/claim-2-theorem-1-2-error-scaling-laws/page.md",
30        "children": []
31      },
32      {
33        "slug": "claim-3-theorem-1-2-matching-bounds",
34        "title": "The derived upper bounds match corresponding lower bounds up to polylogarithmic factors, showing the Newton step is provably more accurate than influence functions whenever d ≫ k (Theorem 1.2).",
35        "file": "pages/claim-3-theorem-1-2-matching-bounds/page.md",
36        "children": []
37      },
38      {
39        "slug": "claim-4-theorem-1-6-1-7-rif-drif",
40        "title": "Rescaled and doubly-rescaled influence functions (RIF and DRIF) are shown to match the Newton step's Õ(kd/n²) average-case error rate while preserving the additivity property that plain influence functions have but the Newton step lacks (Theorem 1.6, Theorem 1.7).",
41        "file": "pages/claim-4-theorem-1-6-1-7-rif-drif/page.md",
42        "children": []
43      },
44      {
45        "slug": "claim-5-remove-lambda-dependence",
46        "title": "The paper's revised bounds remove the problematic 1/λ³ dependence on the regularization coefficient λ present in prior influence-function error analyses (Section 1, Discussion of prior bounds).",
47        "file": "pages/claim-5-remove-lambda-dependence/page.md",
48        "children": []
49      },
50      {
51        "slug": "executive-summary",
52        "title": "Executive summary",
53        "file": "pages/executive-summary/page.md",
54        "children": []
55      },
56      {
57        "slug": "conclusion",
58        "title": "Conclusion",
59        "file": "pages/conclusion/page.md",
60        "children": []
61      }
62    ]
63  },
64  "agent_view_tokens": 5480,
65  "revision": "1784922116190000500"
66}