SOLARIS: a predictive latent cache for serving a recommendation foundation model¶

Fidelity: core mechanism. Executes the future-pair predictor, asynchronous cache, and fallback.

Paper information¶

Field	Details
Paper	arXiv 2604.12110
Organization	Meta
First public release	2026-04-13 (arXiv v1)
Authors' code	The paper did not supply official or author code when checked on 2026-08-09.
Adapter / method	`solaris`
Local reproduction code	`src/auto_research/reproductions/solaris/`

Original paper¶

Motivation and method¶

A large recommender cannot compute latents for every user–item pair at request time. SOLARIS predicts which pairs will be accessed, asynchronously precomputes their latents with a foundation model, and caches them. Requests use cached latents on hits and fall back on misses.

flowchart LR
  A["Future-pair predictor"] --> B["Async foundation model"]
  B --> C["Latent cache"]
  D["Online request"] --> C
  C --> E["Lightweight ranker"]
  D --> F["Fallback"]

Key figure from the paper¶

Figure 1: Overview of the paper's method and key components. From the original paper; copyright remains with the authors.

Core equations¶

\[ \mathcal C_t=\{(u,i):p_\phi(i\mid u,t+\Delta)>\tau\},\qquad s(u,i)=f(e_u,e_i,z_{u,i}^{\rm cache}). \]

Paper results¶

The paper reports full-traffic deployment and top-line revenue +0.67%.

Local reproduction¶

The pair predictor learns from the next three interactions. Only high-confidence pairs are cached; the run measures cache coverage and fallback behavior.

Local comparison: Request-time ranking is the baseline; the predictive latent cache is the treatment. NDCG@10 rises from 0.03540 to 0.09118, or +157.58% relative; see metrics/movielens-100k-seed42.json.

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