Graph Dataset Pruning Formal Evidence

This offline Space presents accepted formal and bounded finite evidence. Paper-reported training measurements remain unavailable.

Target claims

• The paper casts dataset pruning as a graph problem with node weights for intrinsic importance and edge weights for extrinsic diversity/interaction, yielding a Maximum Weight Clique formulation (Section 3.3).
• Under mild conditions, the unified objective becomes submodular and admits a greedy approximation guarantee (Section 3.6; Appendix F).

Pinned paper: arxiv:2606.12913v2
Evidence source revision: f84ea8cd366a28a4fba2b676dd3e033c1350d4d5