Alibaba paper compiles real-time supply constraints into generative retrieval, cutting filtered merchants 170→29
_reachsumit · x · 2026-11-06
An Alibaba team released an arXiv paper on generative retrieval under real-time supply constraints in local retail and delivery.
Problem: generative retrieval predicts semantic IDs for entities, but in local commerce candidates must be serviceable at request time — location, weather, and traffic reshape supply, so a catalog-valid ID may be infeasible; post-generation filtering wastes beam capacity and inference compute.
RTP-LLM pipeline:
- Compiles eligible residual-quantization identifiers into a compressed sparse row prefix index, overlapping CPU compilation with model prefill
- Shares the uploaded index across beams; legal expansions retain original-model log probabilities, constraint states follow selected beam parents
- Formalizes snapshot-relative identifier validity, distinct from entity-level eligibility and freshness
Results: in a matched-beam daily-constraint case, filtered merchants drop from 170 of 299 candidates to 29. The system generates three-token merchant identifiers.
More from Research
- ExPert paper inverts perturbation models to trace cell states back to causal causes — anshulkundaje · 2026-11-06
- Follow-up: gradient flow lets residual states encode information for far-future predictions — mlpowered · 2026-11-06
- Gautam Kamath on AI review: fun science experiment, but false positives can mislead people at scale — thegautamkamath · 2026-11-06
- UCLA's FlowPilot uses human preferences to cut sidewalk robot interventions by 40% — zhoubolei · 2026-11-06
- diffTLM: NeurIPS 2026 Spotlight differentiable vocal tract simulator reconstructs shape from voice — sainingxie · 2026-11-06
- AI model learns thousands of words from one child's audio — only with nightly replay, Science reports — jpillowtime · 2026-11-06