DOS: Dual-Stream Orthogonal Semantic IDs¶

Collaborative/semantic streams, orthogonal rotation, and residual quantization execute.

Paper information¶

Field	Value
Paper	arXiv 2602.04460
Affiliations	Meituan
First public release	2026-02-04 (arXiv v1)
Authors' code	No official author code linked as of 2026-08-09.
Adapter / method	`dos`
Local code	`src/auto_research/reproductions/dos/`

Original paper¶

Background and contribution¶

Content-only codebooks may not match generative recommendation. Align behavior and item semantics, then separate primary and secondary signals with constrained orthogonal rotation before residual quantization.

flowchart LR
  H[Behavior stream] --> A[Alignment]
  I[Item semantics] --> A --> O[Orthogonal rotation] --> R[Residual codebooks] --> G[Generation]

Key figure from the paper¶

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

Core mechanism¶

\[ X_{\mathrm{orth}}=XW,\qquad \mathcal L_{\mathrm{orth}}=\lVert W^\top W-I\rVert_F^2,\qquad r_{\ell+1}=r_\ell-e_{c_\ell}. \]

Paper results¶

A one-week trial on 30% of Meituan traffic reported +1.15% revenue and deployment to hundreds of millions.

Local reproduction¶

Compare unrotated content/collaborative ensemble and Procrustes alignment with three-level ORQ: NDCG@10 0.03540→0.03938 (+11.26%). Rotation error is 6.67e-15; results.

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

Reproduction limits¶

Public genres and co-occurrences substitute for private LLM item vectors and the generator; not comparable with revenue.