Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Uncertainty-driven Embedding Convolution
Sungjun Lim, Kangjun Noh, Youngjun Choi, Heeyoung Lee, Kyungwoo Song
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 |
|---|---|---|---|
| replace uniform ensemble with UEC uncertainty-driven convolution | BGE/E5/GTE embeddings · MMTEB retrieval avg | nDCG@10: 77.76 -> 79.58 | Improved |
| add UEC to embedding ensemble | BGE/E5/GTE · MMTEB retrieval avg | AUC@10: 60.30 -> 67.61 vs uniform ensemble | Improved |
| replace uniform ensemble with UEC | BGE/E5/GTE · MMTEB classification avg | accuracy: 66.79 -> 68.89 | Improved |
| replace uniform ensemble with UEC | BGE/E5/GTE · MMTEB STS avg | Spearman: 75.34 -> 76.49 | Improved |
| uniform and weighted ensembling of deterministic embeddings | BGE/E5/GTE · MMTEB classification (Intent) | accuracy: uniform ensemble 68.78 vs best individual 73.46 | Hurt |
| convert deterministic embeddings to probabilistic via last-layer Laplace approximation | E5, BGE, GTE · MS MARCO | ECE: 0.063 -> 0.036 (E5) | Improved |
| remove uncertainty-aware similarity (ablation) | UEC · MIRACL subset | AUC@10: 91.04 -> 82.48 | Hurt |
| add uncertainty-driven convolution coefficients (remove: use uniform) (ablation) | UEC · MIRACL subset | nDCG@10: 48.45 -> 59.65 | Improved |
| add uncertainty-driven convolution coefficients (ablation) | UEC · MIRACL subset | AUC@10: 10.30 -> 91.04 | Improved |
| add uncertainty-aware similarity and convolution together (ablation) | UEC · MIRACL subset | nDCG@10: 46.78 -> 59.65 | Improved |
| use UEC over a strong single multilingual model | GTE-MB baseline · MMTEB retrieval avg | nDCG@10: 77.33 -> 79.58 | Improved |
| UEC computational overhead | UEC similarity · efficiency | relative time: 1.000 -> 1.006 | No clear effect |
Closest papers by meaning
- Exploiting the Asymmetric Uncertainty Structure of Pre-trained VLMs on the Unit Hypersphere (NeurIPS 2025 · similarity 0.88)
- Repurposing Language Models into Embedding Models: Finding the Compute-Optimal Recipe (NeurIPS 2024 · similarity 0.86)
- Probabilistic Language-Image Pre-Training (ICLR 2025 · similarity 0.85)
- NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models (ICLR 2025 · similarity 0.85)
- Layer by Layer: Uncovering Hidden Representations in Language Models (ICML 2025 · similarity 0.84)
- Adaptive Layer Sparsity for Large Language Models via Activation Correlation Assessment (NeurIPS 2024 · 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.
