FIXLIP: Explaining Similarity in Vision-Language Encoders with Weighted Banzhaf Interactions

2025-11-18 · NeurIPS 2025 · anchor · artifact

Note

date and anchor point to v2, the version held in the project and verified against; the camera-ready title page drops the FIXLIP prefix. Tables 1 and 2 disagree about CLIP ViT-B/32 on the three-object pointing game, 0.83 against 0.82; FX-001 and FX-002 each quote the table they rest on, so the two findings differ by one hundredth on that cell

Findings

• FX-001Pairwise Banzhaf interactions explain CLIP similarity more faithfully than single-score methods
• FX-002FIXLIP gives SigLIP-2 higher pointing-game recognition than CLIP at ViT-B/32 and ViT-B/16
• FX-003FIXLIP's strongest interaction in one CLIP example links doll to an image patch reading dollar