Snap's SetMIR: Multi-Interest Retrieval via Set Prediction
_reachsumit · x · 2026-09-01
Snap presents SetMIR, a multi-interest retrieval method treating retrieval as set prediction.
- Mechanism: Uses a transformer to encode user history and K learnable queries to decode a set of user interests, each with a retrieval embedding and presence score.
- Optimization: Uses Hungarian matching during training to ensure distinct interests; at serving, uses presence scores and NMS to issue only active, non-redundant ANN queries.
- Impact: On Dynamic Product Ads (DPA) data, SetMIR outperforms baselines while issuing 33% fewer queries, boosting CTR by 44% and CVR by 5% in production.
More from Research
- Seoul National University Releases MineAmongUs: VLM Agents Learn to Lie in Embodied Social Settings — SeoulNatlUniv · 2026-09-01
- Naver Proposes Verification-Aware Training to Boost Speculative Decoding Draft Models — naver-ai · 2026-09-01
- Tsinghua researchers break 41-year record, prove Dijkstra is not optimal — jedisct1 · 2026-09-01
- Discussion: RL instills model behaviors independent of system prompts — voooooogel · 2026-09-01
- Abliteration technique removes model refusals while keeping coding/cyber capabilities, sparking debate — aryaman2020 · 2026-09-01
- Explanation of Denoising Diffusion Models and Score Matching — ariG23498 · 2026-09-01