Task-Conditional Flow Matching for Balanced Multilingual Text Embedding Adaptation
TCFM adapts multilingual embeddings by changing the training objective to match the task.
The paper applies Flow Matching only to translation tasks, while using different objectives for retrieval, classification, and pair-classification. It adds teacher-guided representation preservation and a three-stage curriculum to keep adaptation stable. On the Indic Massive Text Embedding Benchmark, the authors report a new state of the art across multilingual tasks and say the method generalizes across embedding model families. Code and datasets are planned for release upon acceptance. HF Daily Papers' note
The paper applies Flow Matching only to translation tasks, while using different objectives for retrieval, classification, and pair-classification. It adds teacher-guided representation preservation and a three-stage curriculum to keep adaptation stable. On the Indic Massive Text Embedding Benchmark, the authors report a new state of the art across multilingual tasks and say the method generalizes across embedding model families. Code and datasets are planned for release upon acceptance. HF Daily Papers' note
score 5