NetEase — The Agent That Audits Its Own Recommender

A deep dive into AutoLR, the agentic harness NetEase built to run the entire lifecycle of improving a production recommender — from proposal generation through offline evaluation to online A/B testing and final Launch Review sign-off — for the feed and immersive-video surfaces of its gaming-community app DASHEN. We cover the business problem (a multi-day, manually-coordinated research-and-engineering loop standing between an idea and a shipped model change), the technical approach (a multi-expert proposal council that debates and adversarially reviews candidates, a deterministic evidence-weighted selector instead of a classic exploration/exploitation bandit, and a hard separation between LLM-generated proposals and machine-validated state transitions), and the realized impact: nine Launch-Review-approved production launches and a system that ran at roughly RMB 3–4 of LLM cost per iteration. We also dig into the paper's most unusual contribution — a self-audit that surfaces a named reliability failure mode, the "KEEP ratchet," where a noise-driven promotion can quietly corrupt the experiment baseline for everyone downstream. Source article: "AutoLR: Automating the Path from Research to Launch Review in Industrial Recommender Systems" — NetEase / arXiv, https://arxiv.org/abs/2609.04871 (published 2026-09).