Kuaishou's AIR framework boosts recommendations 400x with LLM semantics
400x faster LLM reasoning for cross-domain recommendations with real GMV gains
Cross-domain recommendation — using content engagement to predict e-commerce purchases — is notoriously hard due to semantic gaps and noisy, massive user sequences. Large language models (LLMs) offer powerful reasoning but are too slow for real-time systems. Enter AIR (Atomic Intent Reasoning), a new framework from Kuaishou researchers and co-authors that brings LLM semantic capabilities to industrial cross-domain recommendations without the latency penalty.
AIR’s key innovation is splitting the pipeline: LLM inference happens offline, generating atomic intent representations. Online, the system rapidly retrieves and composes these intents based on the user’s real-time context. The result is a 400x inference speedup over direct LLM application while preserving semantic quality. Validated on multiple public datasets, AIR achieves state-of-the-art cross-domain recommendation performance. In live A/B tests on Kuaishou’s massive e-commerce platform, it delivered a +3.446% lift in GMV (gross merchandise value). This makes AIR one of the first practical, high-impact LLM-powered cross-domain recommenders ready for large-scale deployment.
- AIR achieves ~400x inference acceleration by moving LLM reasoning offline and using efficient retrieval/composition online.
- Live A/B testing on Kuaishou E-commerce showed a +3.446% GMV increase, validating real-world commercial value.
- The framework sets new state-of-the-art results on multiple public cross-domain recommendation benchmarks.
Why It Matters
AIR shows LLMs can power real-time recommendations at scale, directly increasing revenue without sacrificing latency.