SPARSEUP paper: training recipe for learned sparse retrievers hits 56.4 nDCG@10 on BEIR-13 under 150M params
antoine_chaffin · x · 2026-11-09
Linkup's research team released SPARSEUP on arXiv, a recipe for training Splade-like sparse encoders on modern backbones. Key contributions: analyzing logit distribution's impact on training stability and sparsity, a per-position fan-out hard cap on vocabulary dimensions, and a vocabulary merging strategy that removes redundant case/word-boundary variants from cased BPE backbones. Applied to LightOn's LateOn backbone, SPARSEUP reaches 56.4 average nDCG@10 on BEIR-13 — the highest known among public vocabulary-based sparse encoders below 150M parameters, with extended comparisons against DenseOn, LateOn and SlideOn.
More from Research
- Formulate a problem as alternating convex optimization and solvers follow naturally — caglar_ee · 2026-11-09
- GitHub study: engineer productivity rises 2.8% per 10% growth of local tech cluster — lemire · 2026-11-09
- 5 PEFT techniques explained: what LoRA, VeRA, Delta-LoRA actually do to model weights — techNmak · 2026-11-09
- Guava distills GPT-5.4 into a 4B robot agent with just 2,000 simulated trajectories — jbhuang0604 · 2026-11-09
- Dex-One2Many learns dexterous manipulation from one human video, 95%+ success across 4 hands — jbhuang0604 · 2026-11-09
- μ0 world model learns robot physics from video only, beats π0 by 5 points on RoboCasa365 — jbhuang0604 · 2026-11-09