Kuaishou's Taiji Boosts Ad Revenue 3.3% for 400 Million Daily Users
Kuaishou's Taiji framework solves the LLM-recommender alignment gap with a new RL algorithm, now live for 400M+ daily users with a 3.3% revenue lift.
Takeaways
- Live at scale: Taiji is deployed on Kuaishou's ad platform since May 2026, serving 400M+ daily users with a 3.3% revenue lift.
- Core problem solved: Bridging the gap between LLM semantic space and recommendation ID space during post-training (SFT + RL).
- POPO algorithm: Pareto Optimal Policy Optimization dynamically reweights LLM semantic rewards vs. collaborative ID rewards, replacing static hand-tuned weights.
- Reverse-engineered CoT: Uses ground-truth user-item outcomes to distill domain-specific chain-of-thought data from QwQ-32B, then filters it with perplexity-based rejection sampling before fine-tuning DeepSeek-R1-7B.
- 2.83% ADVV + 3.30% revenue: Measured via online A/B test on Kuaishou's advertising platform, significant given tens-of-millions revenue base.
- No public release: Production system paper only; no code or model weights are publicly available.
Getting a large language model to make better product recommendations sounds straightforward, but the engineering reality is brutal. LLMs think in rich, open-ended semantic space, while industrial recommenders live and die by tightly indexed item IDs tied to billions of user interactions. Taiji, a new framework from Kuaishou's engineering team, is the most production-validated attempt yet to bridge that gap, and it's already running at scale.
Two worlds that don't naturally talk to each other
Traditional ID-based recommendation models often struggle with semantic understanding, cold-start problems, and cross-domain generalization. That's the motivation for the LLM4Rec wave: inject world knowledge and reasoning into the recommendation pipeline. LLMs offer the potential to introduce extensive world knowledge, robust reasoning capabilities, and high-quality textual generation into the recommendation pipeline.
But the integration is messy. While LLM-based methods demonstrate promise in explainability, their performance on traditional metrics like CTR often falls short of specialized ID-based models, because user behavior is heavily influenced by implicit feedback and specific context rather than the explicit semantic logic found in natural language. The result is a persistent tension: LLM semantic knowledge and collaborative ID signals are both valuable, but they pull in different directions during training.
Aligning the LLM's semantic space with the recommender's ID space via post-training remains challenging. Existing LLM4Rec paradigms are bottlenecked by two main issues: the difficulty of measuring and improving chain-of-thought (CoT) quality in open-domain recommendation during SFT, and the neglect of the trade-off between LLM semantic rewards and recommendation preference rewards during RL alignment.
What Taiji actually does
Taiji (named after the Taiji/Yin-Yang diagram to symbolize the dynamic unity and mutual reinforcement between LLMs and recommender systems) is an industrial-scale LLM-as-Enhancer comprising four key modules: Data Construction, Reasoning Activation, LLM-Recommendation Collaboration, and Online Ranking. Each module targets a specific failure mode in the standard post-training pipeline.