OneFeed unifies search and feed recommendation with generative AI

Modern platforms often run separate models for feed recommendation (which captures implicit browsing interests) and search (which relies on explicit queries). This fragmentation leads to incomplete user profiles and missed opportunities. OneFeed, proposed by Guo Xun, tackles this by introducing a unified generative architecture that jointly models feed content enhancement and query generation. The framework encodes heterogeneous user behavior sequences with a shared encoder, then uses two specialized generative heads: a Feed SemanticID Generator that produces content semantic IDs for recommendation retrieval, and an Intent Query Generator that produces natural-language queries for search-based candidate retrieval. To align these two modalities, OneFeed introduces a SID-Query alignment objective that learns a shared semantic space between content semantic IDs and query representations.

Beyond the core architecture, OneFeed features a closed-loop self-enhancement paradigm that leverages implicit user feedback from both generated content and search-retrieved results to iteratively improve both generation tasks. The paper provides detailed experimental protocols using public recommendation datasets with weakly supervised query construction, along with comprehensive evaluation metrics and expected performance estimates grounded in known baselines. A minimal local prototype validates the pipeline's executability. By treating search and recommendation as two sides of the same coin, OneFeed offers a practical, extensible direction for building more coherent, user-aware systems that can better anticipate needs and reduce redundancy in personalization infrastructure.