Reproduce “Bridging Short Videos and Live Streams: Reasoning-Guided Multimodal LLMs for Cross-Domain Representation Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Le Zhang, Xiaolan Zhu, Yuchen Wang, Shilong Kang, Jiaqi Xue, Xiaoyu Zhang, Xiang Chen, Yalong Guan and 4 more
RGCD-Rep combines structured multimodal reasoning from a frozen 35B MLLM with behavior-grounded cross-domain representation learning to transfer short-video interests to live-stream recommendations. A distilled 2B student encoder and transferable–residual embedding design improve offline retrieval over UniEmbedding, reaching HR@10 0.0254 and NDCG@10 0.0127, while online tests report gains in exposure, clicks, room entry, and follows.
Key insights
- Semantic similarity alone is insufficient for cross-domain recommendation; grounding MLLM reasoning in co-occurrence and engagement behavior helps identify relationships that are actually useful for transfer.
- The pair-level alignment stage contributes more than semantic initialization alone, and updating the student during representation learning outperforms using frozen multimodal embeddings.
- Separating transferable and residual information provides a way to retain useful domain-specific content while reducing alignment of noisy or weakly transferable pairs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.