Juwon Song1, Woohyeong Kim2, Kyeongbo Kong2
1LG Electronics 2Pusan National University
| Project page | arXiv: coming soon | Code: coming soon |

SFPruner consists of three stages:
Semantic guidance
Estimate base token importance using instruction relevance and visual saliency.
Semantics-Guided Ridge Leverage Score (SG-RLS)
Suppress globally redundant covariance directions by reweighting token scores with ridge leverage-based structural information.
Ranking-Based Directional Masking
Resolve residual pairwise redundancy by allowing higher-scoring tokens to suppress similar lower-scoring alternatives through a single parallel masking operation.
This design avoids the iterative dependency of subset-optimization methods such as DPP-, diversity-, or graph-based selection, enabling direct Top-K pruning with stable selection latency.