Netflix reports an LLM-backed ranker beating its production ranker in an A/B test

Sources: paper, HN submission, discussion
Summary: The paper frames GenRec as Netflix exploring a transition from a discriminative ranker carrying thousands of engineered features to an LLM-backed ranker driven by verbalized user histories and context, and covers input verbalization, post-training data construction, reward integration, and a cost-constrained prefill-only serving design. Netflix reports a large-scale A/B test against the current production ranker in which GenRec achieved statistically significant offline and online gains while trained on substantially fewer labelled examples and input signals. The abstract does not state that GenRec is deployed. The paper was submitted 2026-08-10 and last revised 2026-08-21, so it appears here on today's Hacker News discussion rather than as a new release.
Why it matters: This is a deployment report with an A/B result against a production baseline rather than a benchmark claim, and the authors frame the change as moving from feature engineering to context engineering and from bespoke architectures to shared foundation backbones.