PUMA Sparsifies Multimodal Embeddings, Cuts Storage 16x, Speeds Up Retrieval 25x
_reachsumit · x · 2026-08-27
PUMA introduces a post-hoc sparsification recipe using a TopK autoencoder to convert frozen multimodal embeddings into compact sparse codes without retraining the backbone. The method preserves dense dot-product geometry during pretraining before fine-tuning the sparse encoder for retrieval. Evaluations on Qwen3-VL-Embedding-2B show that PUMA matches or outperforms dense retrieval on four out of five benchmarks, reducing vector storage by 8-16x and achieving up to 25x faster exact dense scoring on larger candidate pools.
More from Infra
- Next-gen GraalVM benchmarks: 4x faster startup, 2x less memory, no build step needed — RealGeneKim · 2026-10-15
- How Microsoft optimized Windows for Nvidia's Arm-based RTX Spark AI laptops — tomwarren · 2026-10-15
- How Microsoft optimized Windows for Nvidia's Arm-based RTX Spark laptops — The Verge AI · 2026-10-15
- Global smartphone shipments plunge 22% in worst quarterly decline on record as AI demand fuels memory costs — Polymarket · 2026-10-15
- SK Hynix drops 11% again, now below 3x FY2027 earnings as AI memory selloff deepens — JOBhakdi · 2026-10-15
- ASML Sells Zero Lithography Machines into Europe as US Hosts 75%+ of AI Compute — NinaDSchick · 2026-10-15