{ "schema_version": "1.0", "title": "Reproduction: Finding Most Influential Sets", "emoji": "🔍", "space_id": "snaykey/repro-most-influential-sets", "paper": { "arxiv_id": "2606.05919", "openreview_id": "ghd0zmtpB9" }, "tags": [ "icml2026-repro", "paper-ghd0zmtpB9" ], "updated_at": "2026-07-29T12:00:00+00:00", "root": { "slug": "index", "title": "Reproduction: Finding Most Influential Sets", "children": [ { "slug": "executive-summary", "title": "Executive summary", "children": [] }, { "slug": "claim-1-linear-fractional-reduction", "title": "The Most Influential Sets (MIS) problem for estimands with linear-fractional leave-set-out effects is reduced to a one-parameter sequence of top-k selections solved via Dinkelbach's method (Section 3.1, Algorithm 1)", "children": [] }, { "slug": "claim-2-theorem-1-complexity", "title": "Algorithm 1 costs O(n) per iteration and provably terminates in at most M+1 ratio updates, where M is the number of distinct achievable ratio values, returning a globally optimal set (Theorem 1)", "children": [] }, { "slug": "claim-3-theorem-2-consistency", "title": "Under a uniform-denominator and generated-score-stability assumption, the empirical selection objective uniformly approximates the oracle orthogonal-score objective with o_p(1) error, yielding selection consistency given a separation condition on the ratio gap (Theorem 2)", "children": [] }, { "slug": "claim-4-runtime-scale", "title": "At n=10^6 and k=10^5, the algorithm's median wall-clock runtime is below 200ms, converging in a median of three ratio-update iterations (Section 4.1.2)", "children": [] }, { "slug": "claim-5-algorithm-2-warmstart", "title": "A warm-start variant (Algorithm 2) traces the most influential set across k=1,...,K by initializing each successive top-k problem from the previous solution (Section 3.2)", "children": [] }, { "slug": "conclusion", "title": "Conclusion", "children": [] } ] }, "revision": 1 }