RQ-GMM: Probabilistic residual quantization for multimodal semantics¶

Core public-data mechanism, not the production stack.

Paper information¶

Field	Details
Paper	arXiv 2602.12593
Institution	Tencent.
First public release	2026-02-13 (arXiv v1)
Author code	No public author code found as of 2026-09-05.
Local adapter or method key	`rq-gmm`
Local reproduction code	`src/auto_research/reproductions/rq_gmm/`

Paper method and evidence¶

Quantize image/text embeddings in residual stages with Gaussian-mixture posterior soft assignments to improve codebook utilization relative to hard clustering. Train semantic discretization separately from CTR preference modeling.

flowchart LR
  E[Image and text embeddings] --> G[Mixture posterior]
  G --> Q[Soft residual code]
  Q --> R[Subtract and repeat]

Original paper figure¶

The original figure shows residual quantization and preference modeling. From the original paper; copyright remains with the authors.

Core equation¶

\[ q^{(l)}=\sum_k p(k\mid r^{(l)})\mu_k^{(l)},\qquad r^{(l+1)}=r^{(l)}-q^{(l)}. \]

Reported results¶

Large online A/B reports Advertiser Value +1.502%. This is only author evidence.

Local reproduction¶

Same MovieLens-100K full-catalog protocol: NDCG@10 0.05401→0.04685 (-13.25%). The public proxy cannot predict online lift. Three-seed measurements, mean, SD and 95% CI.

auto-research reproduce --paper rq-gmm --dataset-dir data --seeds 42,43,44

Fixed public subset and auditable proxy, not private logs, production foundation model, traffic splitting or serving.