GALA: Generative Aligned Learning for Adaptive Multimodal Representation in the Taobao Shangou Recommender System¶

Core mechanism reproduction. Paper conclusions, local public-data measurements and unreproduced scope are separate.

Paper information¶

Field	Details
Paper	arXiv 2607.29213
Institution	Rajax Network Technology / Taobao Shangou / Alibaba
First public release	2026-07-31 (arXiv v1)
Author code	No official author code in the paper when checked on 2026-08-24.
Local adapter or method key	`gala`
Local reproduction code	`src/auto_research/reproductions/gala/`

Original paper summary¶

Background and main changes¶

Pretrain multimodal representations on query–image–text triples, align them with GRPO and behavioral rewards, then fuse multimodal and ID representations with an adaptive gate.

flowchart LR
 A["Triple pretraining"] --> B["GRPO behavioral alignment"] --> C["ID and multimodal gating"]

Key figure from the paper¶

Figure 1: Core architecture, modules and connections. From the paper; copyright belongs to the authors. Click for the source.

Core equations¶

\[ z=g\odot z_{mm}+(1-g)\odot z_{id},\quad g=\sigma(W[z_{mm};z_{id}]) \]

Offline and online results¶

The primary online result is order volume +0.55% on a randomized traffic split. Offline tables and online metrics use private or paper-specific protocols, so they cannot be compared directly with MovieLens below.

Local reproduction¶

Comparison: Shared transition + content scorer versus adding only gala on identical users, items, full-catalog candidates and seed. Relative NDCG@10: -5.60%.

MovieLens-100K, 220 users / 360 items, seed 42: NDCG@10 0.0540 → 0.0510 (-5.60%) and Hit@10 0.1091 → 0.1045. Validation alone selects the blend weight; test does not tune it.

auto-research reproduce --paper gala --dataset-dir data --seed 42

See stable metrics.

Scope¶

The mechanism runs on fixed full-catalog MovieLens-100K with identical splits and seed. Company features, production foundation model and traffic are not available; A/B figures are from the paper only. The implementation has distinct model state and scoring, not an alias for a generic weighted formula. See the adapter's omitted_core_components. Local relative changes are not paper A/B lift.