ROCS: Share recommendation compute across candidates in a request¶

Core mechanism: one request encoding, late candidate interaction, and batch scoring.

Paper information¶

Field	Details
Paper	arXiv 2607.27744
Institution	Meta AI.
First public release	2026-07-30 (arXiv v1)
Author code	Author IKBO code.
Local adapter or method key	`rocs`
Local reproduction code	`src/auto_research/reproductions/rocs/`

Paper method and evidence¶

User-side features are shared across many candidates. Generalized Layer Masking isolates candidates, Deep Cross Attention defers candidate interaction, and IKBO broadcasts the request representation inside the GPU kernel.

flowchart LR
  U[Request features] --> E[Encode once]
  C[Candidate batch] --> D[Late cross-attention interaction]
  E --> D
  D --> B[Batch scoring]

Original paper figure¶

Original architecture diagram showing request compute sharing. From the original paper; copyright remains with the authors.

Core equation¶

\[ h_u=E(x_u),\qquad s(u,i)=D(h_u,e_i),\quad i\in\mathcal C_u, \]

\(E(x_u)\) is evaluated once per request.

Reported results¶

Meta deployed this across ads and organic content and across retrieval and ranking. Retrieval achieved up to 3× QPS without quality loss; short-video ranking reports LogLoss −0.5% and QPS +50%.

Local reproduction¶

Relative to per-candidate repeated encoding, the ROCS core gives local NDCG@10 +8.19%. A 128-candidate in-process microbenchmark is 129.58×; it compares repeated vs shared encoding, not production GPU QPS.

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

Fixed metrics. No production GLM, IKBO CUDA kernel, private Meta models, or cluster.