Alibaba proposes TS-GEAR: generative ad retrieval with geo-semantic IDs for delivery search
_reachsumit · x · 2026-11-11
Alibaba researchers published TS-GEAR, a generative retrieval framework for on-demand delivery search ads (e.g., Taobao Shangou), where out-of-range merchants are invalid candidates rather than merely low-ranked. Key designs: Geo-Semantic IDs split merchant identifiers into geographic prefixes and semantic suffixes so reachability becomes a prediction target; Query-Geo Target Attention aggregates user history conditioned on the live query and location; and GMPO adds set-level signals to suppress near-duplicate candidates, addressing the lack of collective-value training signal in autoregressive retrieval.
More from Research
- Skild AI Open-Sources Steel, a Physics Solver Up to 13x Faster Than PhysX With 3x Less GPU Memory — deepakpathak · 2026-11-11
- RoboScript hits 83.4% success on RoboTwin 2.0 with one demo per task, beating trained policies by 20+ points — H-EmbodVis · 2026-11-11
- Alibaba's PhysOPSD uses executable code as privileged context to make video generation physically plausible — AlibabaTongyiLab · 2026-11-11
- mayfer: clean low-rank latents may be the key to continuous online learning — mayfer · 2026-11-11
- PhD student finds lost Mayan city with 6,000+ buildings in old LiDAR dataset — bilawalsidhu · 2026-11-11
- Scaling intelligence: pattern-discovery problems all benefit from test-time scaling — QuanquanGu · 2026-11-11