AI Paper Atlas

Research map / Label-efficient and robust learning / OOD and anomaly detection

Generalizable and Composable Multi-Model Embedding Translation

Beining Yang, Yang Cao

ICML 2026 · Label-efficient and robust learning › OOD and anomaly detection · approach: Robust fine-tuning for OOD generalization

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 changedWhereMeasuredEffect
add directional residual loss (Ldir) to a monolithic translator (ablation)EmbConv-style monolithic translator · FiQA-2018->SciFact mixing, gemini->openaiRecall@100 drop (%): 28.8 -> 3.8Improved
add directional residual loss (Ldir) to a monolithic translator (ablation)EmbConv-style monolithic translator · FiQA-2018->SciFact chaining, gemini->openaiRecall@100 drop (%): 24.6 -> 2.5Improved
add directional residual loss (Ldir) to MoE translator (ablation)H-MoE · FiQA-2018->SciFact mixing, gemini->openaiRecall@100 drop (%): 11.1 -> 2.3Improved
add directional residual loss (Ldir) to monolithic translator (ablation)monolithic translator · FiQA-2018->SciFact OOD, gemini->openaiRecall@100: 83.8 -> 86.7Improved
replace monolithic translator with mixture of local experts (ablation)translator with local loss · FiQA-2018->SciFact OOD, gemini->openaiRecall@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->openaiRecall@100: 89.8 -> 93.4Improved
hierarchical routing instead of flat nearest-centroid routing (ablation)H-MoE · cluster count sweepRecall@100Improved
increase number of clusters (local experts) (ablation)H-MoE · empirical Lipschitz constantLipschitz proxyImproved
scale up training data of nonlinear translatorsEmbConv/Vec2Vec · empirical Lipschitz proxyLipschitz proxyHurt
use H-MoE instead of EmbConv for cross-dataset translationEmbConv · Fever -> other BEIR datasets, 90 pairsRecall@100: average +14.9% over EmbConvImproved
use H-MoE instead of EmbConv under mixingEmbConv · FiQA-2018->SciFact mixingRecall@100 drop: baselines up to 36.6% drop vs H-MoE under 2%Improved
use H-MoE instead of EmbConv under 2-hop chainingEmbConv · FiQA-2018->SciFact chainingRecall@100 drop: 21.1% drop (EmbConv) vs within about 3% (H-MoE)Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy