Snap's GRP: one generative model for retrieval and ranking, +2.56% shares online
_reachsumit · x · 2026-09-30
Snap's GRP v0.1 technical report describes a generative recommendation framework unifying retrieval, ranking, and reward modeling in a single encoder-decoder model. It generates multimodal Semantic IDs, trains a ranking module jointly, and uses the frozen ranker as reward for RL post-training via a new mGRPO objective with reference-anchored margins. Serving optimizations cut end-to-end retrieval latency by 69%. Online: retrieval-only use lifts view time 0.46% and shares 0.77%; with early-ranking bypass and source replacement, +0.82% view time and +2.56% shares with neutral guardrails.
More from Research
- Fleuret: teacher forcing turns time iteration into depth iteration, and GPUs love it — francoisfleuret · 2026-10-24
- Domingos: An attention head is just a fast-weight MLP — pmddomingos · 2026-10-24
- Ross Taylor: studying LLM behavior will increasingly resemble social science — rosstaylor90 · 2026-10-24
- Ross Taylor on RLCR: simplicity is the final achievement, details matter — rosstaylor90 · 2026-10-24
- SenseTime open-sources 27B robot model matching GPT-6-Astra at ~1/50 the cost — rohanpaul_ai · 2026-10-24
- Cohere Labs brings five main track papers to EMNLP 2026 in Budapest — Cohere_Labs · 2026-10-24