MAGIC: training-free optimal-transport compression slashes multi-vector visual document retrieval cost
_reachsumit · x · 2026-09-21
MAGIC is a training-free post-hoc compressor for multi-vector visual document retrieval systems like ColPali. It derives a MaxSim-induced compression surrogate and optimizes it via a two-marginal entropic optimal-transport formulation, prioritizing high-use patches. Across ViDoRe benchmarks and backbones it consistently beats strong baselines, with the largest gains under aggressive compression.
More from Research
- FRAUDSkill Boosts Audio Anti-Fraud Macro-F1 to 73.5% With Frozen Weights — PPSUCTeleantifraudCommunity · 2026-09-21
- TeleAntiFraud 2.0: Chinese Audio Fraud Benchmark Shows F1 Drops to 0.65 on Near-Domain Negatives — PPSUCTeleantifraudCommunity · 2026-09-21
- MoME: Context-Aware Mixture-of-Memory Embeddings Outperform Value Embedding Baselines at Iso-FLOPs — vector-institute · 2026-09-21
- Fitting a performance cone with bootstrapped models to test if AI leaderboard rankings actually hold — PTenigma · 2026-09-21
- AI writing detectors flagged her lab vision post; she asks what we should actually measure — furongh · 2026-09-21
- Human-AI Collaboration Settles Major Open Problem in Multi-Winner Voting Theory — xuanalogue · 2026-09-21