Frozen LVLMs for Micro-Video Recommendation: A Systematic Study of Feature Extraction and Fusion

This project is a research-centric systematic study rather than a production-grade software product. With 0 stars and 6 forks after 100 days, it lacks community traction and serves primarily as a reference for the associated arXiv paper. The core contribution is the empirical evaluation of existing LVLMs (like Video-LLaVA) within recommendation pipelines, specifically testing how to fuse high-dimensional semantic features with collaborative filtering ID embeddings. From a competitive standpoint, the 'moat' is non-existent; the techniques described (feature projection, concatenation, or gated fusion) are standard practitioners' patterns in the industry. The primary risk comes from platform giants (ByteDance, Meta, Google) who already possess proprietary, much larger versions of these pipelines. Frontier labs like OpenAI or Google could easily release 'Video-Embedding-001' APIs that render these extraction strategies obsolete by providing more 'recommendation-ready' latent spaces. While useful for researchers looking for a baseline, it offers no unique data gravity or technical barrier to entry. The 'frozen' nature of the models mentioned is a common cost-saving heuristic in production, but not a novel architectural breakthrough.