Alibaba finds generative recommendation still has hard limits on cold items
_reachsumit · x · 2026-07-24
Alibaba studies why generative recommendation still struggles with cold items
Alibaba explores Semantic-ID generation for generative recommendation under an absolute-time protocol.
- The paper shows that cold-item reachability is bounded by the learned token paths.
- In other words, even when generation is used for recommendation, the model may still have structural limits in surfacing new or rarely seen items.
- The post links to the paper for the full method and analysis.
More from Research
- CORTEX turns local models into an interpretability desktop with logits, attention, and interventions — JayB_Official · 2026-07-24
- IBM Research features work on standardizing AI evals — evijit · 2026-07-24
- Why agent stacks can keep taking valid steps while losing the verified state — Present-Quantity-813 · 2026-07-24
- AREX introduces a recursively self-improving deep research agent with inner and outer loops — _reachsumit · 2026-07-24
- NeurIPS reviewer joke meets Pangram AI’s all-or-nothing scoring screenshot — torchcompiled · 2026-07-24
- Editable text user profiles make recommendations more controllable — _reachsumit · 2026-07-24