GRP v0.1: Progressive end-to-end generative recommendation

简体中文

L1 architecture/objective diagnostic, not Snap online reproduction.

Paper information

Field	Details
Paper	arXiv v1
Institution	Snap Inc.
First public release	2026-09-29 (arXiv v1)
Author code	No author repository found as of 2026-09-30.
Local adapter or method key	grp
Local reproduction code	src/auto_research/reproductions/grp/

Paper method and evidence

Gradually add the unified generator as a retrieval source, replace weaker sources, increase its quota and bypass rankers. Encoder–decoder outputs multimodal SIDs. MHP ranks from detached encoder/SID states and then serves as frozen RL reward.

flowchart LR
  H[Long histories] --> E[Encoder]
  E --> B[Independent decoder blocks]
  B --> S[Semantic IDs]
  E -. stop gradient .-> M[MHP]
  S -. stop gradient .-> M
  M --> R[mGRPO]

Original paper figure

First-page overview of the architecture. From the original paper; copyright remains with the authors.

Core equation

mGRPO keeps clipped policy ratios and adds a one-sided logged-target margin only when the current policy sacrifices target retrieval beyond tolerance. Ranking loss cannot reach the generator.

Reported results

Online retrieval-only comparison: view time +0.46%, shares +0.77%. Bypass plus weak-source replacement: +0.82%, +2.56%, and end-to-end retrieval latency 69% lower. Author industrial P0 evidence, not local measurements.

Local reproduction

Run auto-research reproduce --paper grp for block isolation, candidate score differentiation and finite generation/mGRPO gradients. Same GRPO surrogate without the one-sided guard is the comparison; no capability estimate. Three-seed metrics. No private Snap logs, Qwen3-VL SID tokenizer, production training, CUDA graph serving, inverted index or live traffic. concept_demo / l1_mechanism, not Evolve gain evidence.