Research map / Label-efficient and robust learning / OOD and anomaly detection
Generalizable and Composable Multi-Model Embedding Translation
Beining Yang, Yang Cao
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 directional residual loss (Ldir) to a monolithic translator (ablation) | EmbConv-style monolithic translator · FiQA-2018->SciFact mixing, gemini->openai | Recall@100 drop (%): 28.8 -> 3.8 | Improved |
| add directional residual loss (Ldir) to a monolithic translator (ablation) | EmbConv-style monolithic translator · FiQA-2018->SciFact chaining, gemini->openai | Recall@100 drop (%): 24.6 -> 2.5 | Improved |
| add directional residual loss (Ldir) to MoE translator (ablation) | H-MoE · FiQA-2018->SciFact mixing, gemini->openai | Recall@100 drop (%): 11.1 -> 2.3 | Improved |
| add directional residual loss (Ldir) to monolithic translator (ablation) | monolithic translator · FiQA-2018->SciFact OOD, gemini->openai | Recall@100: 83.8 -> 86.7 | Improved |
| replace monolithic translator with mixture of local experts (ablation) | translator with local loss · FiQA-2018->SciFact OOD, gemini->openai | Recall@100: +6.0% on P1 (86.7 -> 89.8 in table) | Improved |
| add directional loss on top of MoE (ablation) | H-MoE · FiQA-2018->SciFact OOD, gemini->openai | Recall@100: 89.8 -> 93.4 | Improved |
| hierarchical routing instead of flat nearest-centroid routing (ablation) | H-MoE · cluster count sweep | Recall@100 | Improved |
| increase number of clusters (local experts) (ablation) | H-MoE · empirical Lipschitz constant | Lipschitz proxy | Improved |
| scale up training data of nonlinear translators | EmbConv/Vec2Vec · empirical Lipschitz proxy | Lipschitz proxy | Hurt |
| use H-MoE instead of EmbConv for cross-dataset translation | EmbConv · Fever -> other BEIR datasets, 90 pairs | Recall@100: average +14.9% over EmbConv | Improved |
| use H-MoE instead of EmbConv under mixing | EmbConv · FiQA-2018->SciFact mixing | Recall@100 drop: baselines up to 36.6% drop vs H-MoE under 2% | Improved |
| use H-MoE instead of EmbConv under 2-hop chaining | EmbConv · FiQA-2018->SciFact chaining | Recall@100 drop: 21.1% drop (EmbConv) vs within about 3% (H-MoE) | Improved |
Closest papers by meaning
- Scaling Embedding Layers in Language Models (NeurIPS 2025 · similarity 0.86)
- When is an Embedding Model More Promising than Another? (NeurIPS 2024 · similarity 0.85)
- EmbedLLM: Learning Compact Representations of Large Language Models (ICLR 2025 · similarity 0.84)
- DH-Set: Improving Vision-Language Alignment with Diverse and Hybrid Set-Embeddings Learning (CVPR 2025 · similarity 0.84)
- MoME: Mixture-of-Memory Embeddings for Context-Aware Sparse Lookup (NeurIPS 2026 · similarity 0.84)
- Your Mixture-of-Experts LLM Is Secretly an Embedding Model for Free (ICLR 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.
