CS3¶

Core-mechanism reproduction. Cycle revision, cross-tower synchronization, and cascade-teacher auxiliary loss.

Paper information¶

Field	Value
Paper	arXiv 2604.19269
Affiliations	Kuaishou Technology
First public release	2026-04-21 (arXiv v1)
Authors' code	Author code.
Adapter / method	`cs3`
Local code	`src/auto_research/reproductions/cs3/`

Original paper¶

Background and contribution¶

Independent towers enable ANN but limit interactions. CS3 adds cycle revision and knowledge coordination while preserving offline and online deployability.

flowchart LR
  U[User tower] --> R[Cycle self-revision]
  I[Item tower] --> R
  R --> S[Cross-tower sync] --> Y[ANN score]
  T[Cascade teacher] --> Y

Key figure from the paper¶

Original Figure 2: CS3 architecture. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ u'=u+g_u(u,i)\odot\Delta_u,\quad i'=i+g_i(i,u)\odot\Delta_i,\quad L=L_{\rm rank}+\lambda L_{\rm cascade}. \]

Paper results¶

Revenue rose 8.356% / 1.366% / 2.177% across three Kuaishou production settings.

Local reproduction¶

Matched two towers are the baseline. Cycle, cross sync, and teacher loss gave −16.06% relative three-seed NDCG@10, not the online gain. See metrics.

auto-research reproduce --paper cs3

Reproduction limits¶

No 400-million-DAU cache, EMA cross-tower vectors, or production QPS path.