MESH: Scalable retrieval with decoupled heterogeneous content¶

Core grouped manifolds, independent residual amplifiers, global crossing, and RGBC.

Paper information¶

Field	Details
Paper	arXiv 2607.12392
Institution	Pinterest.
First public release	2026-07-14 (arXiv v1)
Author code	No official author code linked by the paper as checked on 2026-08-09.
Local adapter or method key	`mesh`
Local reproduction code	`src/auto_research/reproductions/mesh/`

Paper method and evidence¶

Flat retrieval scaling can let high-frequency evergreen gradients overwhelm fresh/tail signals. Separate user, item, and context subtowers and amplify each via residual crossing before global interaction. RGBC gates user–item affinity with context; serving can parallelize subtowers and cache item features asynchronously.

flowchart LR
  U[User subtower] --> G[Global crossing]
  I[Item subtower] --> G
  C[Context subtower] --> G
  C --> A[Context-gated affinity]
  G --> A

Original paper figure¶

The original figure shows separated subtowers and gated affinity. From the original paper; copyright remains with the authors.

Core equation¶

\[ H_k^l=\operatorname{Norm}(H_k^{l-1}+\Phi(H_k^{l-1},H_k^{l-1})),\qquad H_0=\operatorname{LN}(\sigma(MLP(T_c))\odot(T_u\odot T_i)+T_c). \]

Reported results¶

Pinterest reports fresh-item repins +5.5%, funnel efficiency +55%, retention +0.46%, and system throughput 2.87×.

Local reproduction¶

Flat retrieval baseline versus three amplified towers and RGBC. Validation selects blend 0.1; local Hit@10 +0.00%, NDCG@10 -3.54%. Small Python batch microbenchmark 0.33×, retained as a counterexample rather than compared with TorchScript/GPU parallel serving. Metrics.

auto-research reproduce --paper mesh --seed 42

No complete DHEN or production GPU inter-op; local timings do not validate paper throughput.