Research map / Architectures and efficiency / Pruning and efficient tuning
CSRv2: Unlocking Ultra-Sparse Embeddings
Lixuan Guo, Yifei Wang, Tiansheng Wen, Yifan Wang, Aosong Feng, Bo Chen, Stefanie Jegelka, Chenyu You
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 k-annealing (ablation) | CSR · MTEB (6 task types, avg), k=2, e5-Mistral-7B | average score: 44.17 -> 45.33 | Improved |
| add supervised contrastive loss (ablation) | CSR · MTEB (6 task types, avg), k=2, e5-Mistral-7B | average score: 44.17 -> 49.05 | Improved |
| combine k-annealing and supervision (ablation) | CSR · MTEB (6 task types, avg), k=2, e5-Mistral-7B | average score: 49.05 -> 53.25 | Improved |
| add full backbone finetuning (ablation) | CSRv2-linear · MTEB (6 task types, avg), k=2, e5-Mistral-7B | average score: 53.25 -> 58.34 | Improved |
| add k-annealing (ablation) | CSR · MTEB avg, k=64, e5-Mistral-7B | average score: 66.68 -> 67.35 | Improved |
| add full finetuning (ablation) | CSRv2-linear · MTEB avg, k=64, e5-Mistral-7B | average score: 67.58 -> 68.08 | Improved |
| replace CSR with CSRv2 | CSR · MTEB avg, k=2, Qwen3-Embedding-4B | average score: 36.29 -> 58.53 | Improved |
| replace CSR with CSRv2 at full dimension | CSR · MTEB avg, 4096 active dims, e5-Mistral-7B | average score: 69.70 -> 69.80 | No clear effect |
| use CSRv2 instead of SPLADEv3 at same low budget | SPLADEv3 · MTEB Retrieval (8 datasets), k=2 | nDCG@10: 11.34 -> 37.48 | Improved |
| use CSRv2 instead of MRL at k=2 | MRL · MTEB avg, k=2, e5-Mistral-7B | average score: 33.81 -> 58.38 | Improved |
| k-annealing reduces dead neurons (ablation) | CSR · dead neuron ratio, k=2 | dead neuron ratio: 90% -> about 20% (with supervision) | Improved |
| finetuning backbone effect on dead neurons (ablation) | CSRv2-linear · dead neuron ratio | dead neuron ratio | No clear effect |
Closest papers by meaning
- Beyond Matryoshka: Revisiting Sparse Coding for Adaptive Representation (ICML 2025 · similarity 0.89)
- A Single-Step, Sharpness-Aware Minimization is All You Need to Achieve Efficient and Accurate Sparse Training (NeurIPS 2024 · similarity 0.85)
- NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models (ICLR 2025 · similarity 0.84)
- From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit (NeurIPS 2025 · similarity 0.84)
- Sparse Dimensionality Reduction Revisited (ICML 2024 · similarity 0.84)
- CoreMatching: A Co-adaptive Sparse Inference Framework with Token and Neuron Pruning for Comprehensive Acceleration of Vision-Language Models (ICML 2025 · similarity 0.84)
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.
