CCFormer: Cross-Field Interaction and Hierarchical Sequence Compression¶

Local core-mechanism evidence. Paper outcomes and local results are separate.

Paper information¶

Field	Value
Paper	arXiv 2607.28070
Affiliations	Tencent Platform and Content Group
First public release	2026-07-30 (arXiv v1)
Authors' code	No official author code linked as of 2026-08-09.
Adapter / method	`ccformer`
Local code	`src/auto_research/reproductions/ccformer/`

Original paper¶

Background and contribution¶

Industrial ranking needs field interactions and thousand-event histories. Project ID/content fields separately before cross-attention; widen receptive fields layerwise to compress old history while retaining detailed recent tokens.

flowchart LR
  F[Separate field features] --> C[Cross-field interaction] --> Y[Multitask predictions]
  H[Long history] --> Z[Hierarchical compression] --> M[Token mixing] --> Y

Key figure from the paper¶

Original CCFormer architecture. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ H^{(l+1)}=H^{(l)}+\operatorname{Mix}_l(\operatorname{Compress}_l(H^{(l)})),\qquad Q_f=X_fW_f^Q. \]

Paper results¶

Two Tencent settings, over 1M daily exposed users per arm, two weeks: +3.57% video CTR and up to +1.71% ad revenue. Fully deployed; 2.21× training speed versus HSTU.

Local reproduction¶

Matched-budget GRU versus CCFormer: Hit@10 0.0208→0.0292, NDCG@10 0.01092→0.01337 (+22.44%), encoded tokens 24→12. Metrics.

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

Reproduction limits¶

No 4B private examples, thousand-event production sequences, or parallel multitarget kernels.