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
- ASML CFO says US Match Act restricting DUV sales to China would backfire — pstAsiatech · 2026-10-15
- ASML sold out through 2028 but stock dips 2%; TL20 index up 57% YTD — TiernanRayTech · 2026-10-15
- Bloomberg breakdown: a 1GW Texas AI data center may need nearly $3B for overruns alone — rohanpaul_ai · 2026-10-15
- 7 System Design Algorithms to Know: Merkle Trees to CAP/PACELC — goyalshaliniuk · 2026-10-15
- MiniCPM-V 4.7 runs a 35B MoE multimodal model on phones with just 3GB memory — nevrekaraishwa2 · 2026-10-15
- Sipeed MaixStation product page goes live for edge AI dev — xlltt · 2026-10-15