Trainable [DUMP] Prefix Tokens Clean Up Retrieval Embeddings

Researchers from Korea University and others propose appending trainable [DUMP] prefix tokens during embedding model training to absorb noisy residual signals, improving retrieval performance for both encoders and decoders.

2026-11-06 ~ 2026-11-06 · 2 related posts

• Trainable [DUMP] prefix tokens reduce representational interference in retrieval embeddings — _reachsumit · 2026-11-06
• Trainable [DUMP] Prefix Tokens Reduce Representational Interference in Retrieval Embeddings — nlpai-lab · 2026-11-06