What Drives Intelligent Recommendation and Search? Key Takeaways from Xiaohongshu’s CCF C³ Event

The CCF C³ activity, organized by the CCF CTO Club, took place on March 30 at Xiaohongshu, bringing together more than 40 experts and scholars to discuss "Intelligent Recommendation and Search". The live broadcast across multiple platforms attracted over 16,000 online viewers, marking a record for the series.

Xiaohongshu, a fast‑growing mobile‑first UGC platform with over 260 million monthly active users and 20 million creators, aims to help ordinary people share and discover content. Its recommendation system differs from typical industry solutions by emphasizing equitable distribution rather than pure efficiency, which requires a decentralized traffic‑allocation mechanism and poses significant algorithmic challenges.

During the technical presentation, Vice President Feng Di highlighted four core challenges: multimodal heterogeneous content recommendation, decentralized distribution, interest diversity, and community “break‑out” (cross‑group interaction), along with cost‑control and compute‑optimization concerns. The team leverages billion‑scale multimodal pre‑training to obtain vector representations of images, text, and video, and is moving toward joint modeling of content and user behavior.

To address cold‑start and real‑time relevance, Xiaohongshu upgraded its high‑frequency recommendation pipeline from daily to minute‑level updates, enabling rapid promotion of new notes and niche content. Forgetting strategies are applied to decay stale user interests, preserving diversity and preventing filter bubbles.

Professor Li Chenliang from Wuhan University presented the state‑of‑the‑art in the recall stage of search‑recommendation systems, emphasizing ultra‑low latency, dual‑tower and deep network models, multi‑interest modeling, long‑tail data handling, and the enrichment of external semantic data. He stressed that recall must balance speed with performance, a challenge he described as both technical and artistic.

The roundtable discussion explored how large language models (LLMs) such as ChatGPT may reshape recommendation and search. Experts noted that LLMs could change user search habits and advertising models, enable more personalized and human‑like recommendations, but also raise concerns about evaluation, privacy, data quality, and the need for domain‑specific fine‑tuning.

Concluding remarks projected a future where next‑generation recommendation technology becomes a unified foundation, with data and compute resources reshaping the industry landscape. The panel agreed that recommendation systems may evolve into a core operating system for AI‑driven information delivery.