Improving an industrial recommender is an iterative research‑engineering loop rather than a one‑shot path from idea to deployment. In NetEase’s gaming‑community app DASHEN, algorithm engineers scout promising directions from papers, technical reports, and past production experiments, reproduce or adapt the methods, integrate them into the production codebase, and validate via offline training and evaluation. Candidates that perform well move to online A/B tests, and those showing robust gains are submitted to Launch Review before full‑traffic rollout. Large language models (LLMs) can assist at individual stages, but the lack of a unified orchestration framework keeps the overall workflow heavily human‑driven, especially across multi‑day experimental cycles. To address this, we introduce AutoLR (originally Auto Launch Review), which has been extended into an end‑to‑end research‑to‑launch harness. AutoLR comprises three core mechanisms: (1) a multi‑expert council that debates and adversarially reviews each proposal; (2) a deterministic evidence‑weighted exploration‑exploitation selector that allocates a limited trial budget across candidate directions and re‑ranks them using the council’s feedback; (3) a layered knowledge system that fuses external research, production‑system expertise, and DASHEN‑specific domain knowledge (e.g., game communities, player characteristics, content‑interaction patterns) with posterior evidence from configurations, patches, logs, failures, and offline outcomes. LLM agents handle semantic reasoning and code generation, while deterministic controllers retain authority over execution, metric extraction, guardrails, and persistent state transitions.
Overall, AutoLR structurally combines human expert insight with LLM capabilities to achieve cross‑stage automation, offering a reproducible pathway to boost development efficiency in large‑scale recommender systems. Review