Think When Needed: Dual-LoRA Architecture Skips Unnecessary Chain-of-Thought in Multimodal Retrieval

arXiv·medium signal

Alibaba introduces TWN, a unified multimodal embedding framework with adaptive reasoning that attaches separate reasoning and embedding LoRA adapters to a shared frozen backbone. An adaptive think mechanism selectively invokes chain-of-thought only when it benefits retrieval quality, skipping redundant reasoning for simple inputs that discriminative embeddings already handle well. Gradient detachment at the adapter interface mitigates optimization conflicts while maintaining near-single-model parameter count.