Pruning SlideOn to 64+512 features costs just 0.02 nDCG while cutting FLOPS 3.5x
tomaarsen · x · 2026-10-20
Matryoshka training concentrates useful information in the strongest features. In sentence-transformers you set topk separately for queries (fewer posting lists visited) and documents (fewer weights stored). Capping SlideOn at 64 active query features and 512 document features drops BEIR-15 only from 56.68 to 56.66 while the retrieval FLOPS proxy falls from 30.87 to 8.72—the vector stays 32,768-dim, keeping its largest activations.
Related event: LightOn Open-Sources SOTA Sparse Retrievers SlideOn and mSlideOn(26 posts)→
More from Models
- OpenAI API adds 12-hour prompt caching option alongside 30-minute default — OpenAIDevs · 2026-10-21
- antirez on model upgrades: 4.0→4.1 knowledge transfers unless attention changes — teortaxesTex · 2026-10-21
- DeepSeek reportedly profits at $0.15/$0.60 off-peak pricing — the price Western labs must beat — arthurcolle · 2026-10-21
- SpecterOps Researchers Show Adversarial ML Evading EDR, LLMs Speeding Reverse Engineering — dyn___ · 2026-10-21
- AWS fine-tunes GPT-OSS-20B for multilingual reasoning: base model reasoned in French 0% of the time — AWS ML Blog · 2026-10-21
- Riemann-bench: frontier models jump from under 5% to over 70% in six months, NS and UGC solved — echen · 2026-10-21