Kuaishou's RECAP boosts video recommendations 4.9% with LLM feedback loops
Closed-loop LLM profiles improve recall by 4.9% and app usage by 0.139%...
Short-video platforms like Kuaishou need to continuously update user profiles as new behavior streams arrive, but most LLM-based profile generators operate in an open loop — they summarize past behavior well but aren't trained to improve future recommendations. RECAP closes that loop by treating profile generation as a reinforcement learning problem, using historical implicit feedback to directly optimize the profile for recommendation accuracy.
The framework maintains each user profile as a bounded structured memory, combining LLM-based semantic updates with deterministic lifecycle and capacity controls. To generate profile-targeted feedback, RECAP filters label-consistent behavior pairs using an LLM judge, then trains a dual-tower evaluator whose matching score serves as a GRPO reward. Experiments on real Kuaishou short-video data show RECAP outperforms the base generator with 0.0084 improvement in uAUC and about 4.9% improvement in Recall@2000. A seven-day online A/B test confirmed a statistically significant 0.139% lift in average app usage time per user, demonstrating real-world impact. The paper, accepted at RecSys 2026, also shows that RECAP produces more grounded and user-level abstract profile updates.
- RECAP uses GRPO (group relative policy optimization) to reward LLM profile updates that improve future recommendation accuracy.
- Offline tests on Kuaishou data show a 4.9% boost in Recall@2000 and +0.0084 uAUC over the baseline generator.
- Online A/B test over 7 days yielded a statistically significant 0.139% increase in average user app usage time.
Why It Matters
LLM profiles now learn from their own mistakes — expect smarter, more adaptive personalization in short-video feeds.