TransRetrieval: Compressed Transformers for Cross-Domain Ad Retrieval¶

Core mechanism evidence. Paper and local evaluations are separate.

Paper information¶

Field	Value
Paper	arXiv 2608.25528
Affiliations	Renmin University of China / Taobao & Tmall Group, Alibaba
First public release	2026-08-26 (arXiv v1)
Authors' code	No public author code found as of 2026-08-28.
Adapter / method	`transretrieval`
Local code	`src/auto_research/reproductions/transretrieval/`

Original paper¶

Background and contribution¶

Reduce source noise, target cost, and domain confusion. Aggregate source histories by token norms, compress target tokens to one, and encode domain as position.

flowchart LR
  S[Source history] --> A[Norm-aware aggregation] --> F[Fused query]
  T[Target tokens] --> C[One target token] --> F
  D[Domain position] --> F

Key figure from the paper¶

Original Figure 1: TransRetrieval. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ \bar h_s=\frac{\sum_t\lVert h_t\rVert h_t}{\sum_t\lVert h_t\rVert},\qquad z_t=W_t\bar h_t+e_{domain,pos}. \]

Paper results¶

A one-month, 5%-traffic production A/B reports +2.53% revenue and +1.28% RPM (\(p<0.0001\)).

Local reproduction¶

Against Transformer two-tower, add three mechanisms on MovieLens-1M: Hit@10 0.1063→0.1187 (+11.76%), NDCG@10 0.05429→0.05473 (+0.81%). Metrics.

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

Reproduction limits¶

No 40B interactions, 52M ads, or production ANN. 85% target FLOPs reduction is structural, not measured latency.