CORAL: An LLM-native research loop for production recommendation

简体中文

Bounded proposal, experiment memory, and feedback loop on public data.

Paper information

Field	Details
Paper	Paper or original source
Institution	Meta AI.
First public release	2026-09-02 (arXiv v1)
Author code	No public author code found as of 2026-09-06.
Local adapter or method key	coral
Local reproduction code	src/auto_research/reproductions/coral/

Paper method and evidence

Have an LLM propose structured configurations from objectives, budget, and recent experiments. Validate constraints, execute, and write outcomes into the next round's short memory. This is an iterative loop rather than one-shot tuning.

flowchart LR
  A[Objectives and short memory] --> B[Propose bounded configuration]
  B[Propose bounded configuration] --> C[Validate and execute]
  C[Validate and execute] --> D[Feed observed results into next round]

Original paper figure

The original figure shows production research feedback. From the original paper; copyright remains with the authors.

Core equation

Project \(x_t=\Pi_{\mathcal C}(x_{t-1}+\Delta_t)\) onto \(\mathcal C=\{x\mid x_i\ge 0,\sum_i x_i\le B\}\). Keep only three recent \((x,e)\) observations to avoid stale feedback.

Reported results

Across millions of Meta users, video sessions +0.16%, watch time +0.15%, and low-signal new-user sessions +0.23%, without more serving cost. In another service the second round increased annualized capacity savings by another 44% with neutral engagement.

Local reproduction

Five budget-allocation rounds on MovieLens-100K; memory of three, validation-only blend. Seed 42 equal allocation baseline NDCG@10 0.05401, intervention 0.05021 (-7.02%). All three seeds and traces. Only controller and constraints validated; no private telemetry, production LLM or traffic. Can combine as an Evolve recommendation controller operator.