Research map / Label-efficient and robust learning / Recognition and long-tail
MILCO: Learned Sparse Retrieval Across Languages via a Multilingual Connector
Thong Nguyen, Yibin Lei, Jia-Huei Ju, Eugene Yang, Andrew Yates
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LexEcho source view to English-only view (ablation) | MILCO · MIRACL dev | nDCG@10: 69.4 -> 72.3 | Improved |
| add LexEcho source view (ablation) | MILCO · MLDR | nDCG@10: 69.9 -> 74.4 | Improved |
| add Sparse Alignment Pretraining before contrastive training (ablation) | MILCO · MIRACL dev | nDCG@10: 59.2 -> 72.3 (with SAP, distill, LexEcho; English-only 59.2 -> 69.4) | Improved |
| add sparse contrastive training after alignment (ablation) | MILCO · MIRACL dev | nDCG@10: 54.5 -> 70.1 | Improved |
| add KL distillation from cross-encoder instead of InfoNCE (ablation) | MILCO · MIRACL dev | nDCG@10: 70.1 -> 72.3 | Improved |
| add multilingual connector (map to English lexicon) instead of full multilingual MLM head (ablation) | MILCO · MIRACL dev | nDCG@10: 50.7 -> 59.2 | Improved |
| use sparse alignment instead of dense alignment (with distill) | MILCO · MIRACL dev | nDCG@10: 67.9 -> 72.3 | Improved |
| use MILCO vs Qwen3-Embed-8B on MTEBv2 | MILCO · Multilingual MTEBv2 | nDCG@10 avg: 75.59 vs 66.83 | Hurt |
| post-hoc mass pruning p=10 (ablation) | MILCO · MIRACL English Seismic | nDCG@10: 56.4 -> 56.3 | No clear effect |
| post-hoc mass pruning p=60 (ablation) | MILCO · MIRACL English Seismic | nDCG@10: 56.4 -> 50.3 | Hurt |
| use MILCO shared English lexical space instead of source-language sparse models | MILCO · MKQA | Recall@100: 45.3 -> 76.6 | Improved |
Closest papers by meaning
- FreeRet: MLLMs as Training-Free Retrievers (ICML 2026 · similarity 0.87)
- MMTEB: Massive Multilingual Text Embedding Benchmark (ICLR 2025 · similarity 0.87)
- Can Language Models Actually Retrieve In-Context? Drowning in Documents at Million Token Scale (NeurIPS 2026 · similarity 0.87)
- Benchmarking and Building Long-Context Retrieval Models with LoCo and M2-BERT (ICML 2024 · similarity 0.87)
- Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps (NeurIPS 2026 · similarity 0.87)
- Repurposing Language Models into Embedding Models: Finding the Compute-Optimal Recipe (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
