RecGPT-V3 Technical Report
RecGPT-V3 reports production gains on Taobao while cutting serving resource use by 52.4%.
The paper says the system adds a Memory Hub to keep evolving user memory instead of rebuilding behavior context on every request. It also mixes natural-language tags with Semantic IDs so the model can reason about intent while grounding recommendations in item space. Its latent intent method compresses long rationales into learnable tokens, with a reported 200x reduction in output token cost. In Taobao’s “Guess What You Like” feed, the reported A/B lifts include GMV +3.97%, TC +1.97%, IPV +1.28%, and CTR +1.00%. HF Daily Papers' note
The paper says the system adds a Memory Hub to keep evolving user memory instead of rebuilding behavior context on every request. It also mixes natural-language tags with Semantic IDs so the model can reason about intent while grounding recommendations in item space. Its latent intent method compresses long rationales into learnable tokens, with a reported 200x reduction in output token cost. In Taobao’s “Guess What You Like” feed, the reported A/B lifts include GMV +3.97%, TC +1.97%, IPV +1.28%, and CTR +1.00%. HF Daily Papers' note
score 5