Generative end-to-end ad retrieval at Douyin headlines weekly IR papers roundup
_reachsumit · x · 2026-10-05
Sumit published Vol. 176 of the weekly Information Retrieval papers newsletter, covering 11 papers from ByteDance, Google, Meta, LinkedIn, Snap, and Shopify. Highlights:
- GEAR (ByteDance): a generative end-to-end ad retrieval system for Douyin that jointly trains the item tokenizer, autoregressive generator, and reranker on streaming data. Its BasisVQ tokenizer multiplies the codebook by a learnable approximately orthogonal basis (via Newton-Schulz iterations) to mitigate codebook collapse under distribution shift.
- Other picks include training-free dense retrieval from LLMs with in-context examples, in-graph curriculum weighting for fresh/tail content in short-video recommendation (Google), task-aware embedding subspaces for multi-objective retrieval (LinkedIn), in-context product search for small catalogs (Google), a timestamp-based rotary embedding for sequential recommendation (Shopify), a unified generative model for retrieval, ranking, and reward (Snap), and measuring lifecycle overhead in large-scale recommendation training (Meta).
More from Research
- Xaira unveils AI drug discovery platform: 10x medicines goal, early results on hard GPCR target — BoWang87 · 2026-10-05
- Bergson: open-source library unifies data attribution methods to study LLM generalization and misalignment — zetalyrae · 2026-10-05
- Aeon classic revisited: your brain does not process information and is not a computer — AnnaCiaunica · 2026-10-05
- The top 50 AI researchers ranked by citations, Attention authors all on the list — ksprdk · 2026-10-05
- Richard Ngo: the hard part of Lobian cooperation is deciding to do it, not hardcoding it — RichardMCNgo · 2026-10-05
- ML educator: learning calculus was a waste of time — and my best career decision — TivadarDanka · 2026-10-05