ZoRRO, a zero-weight training-free news recommender, achieved click-through rates comparable to a state-of-the-art deep learning model in a live online A/B test at Danish publisher JP/Politikens Media Group, while running 600× faster at inference, as reported in a paper accepted to SIGIR '26, held July 20–24, 2026, in Melbourne.
ZoRRO combines article recency with representation-based similarity. Articles are encoded at indexing time using one-hot category encodings or pre-trained vectors such as BERT; ZoRRO does not carry any trained weights. At serving time, it scores candidates by computing similarity between user history representations and fresh article vectors, then applies a recency decay. The system adapts instantly to new articles and users, bypassing the cold-start latency that typically necessitates frequent batch retraining or expensive real-time feature updates in news recommender pipelines. The serving stack is CPU-only, requiring no GPU for inference, no monitoring for embedding drift, and no gradient updates running in the background.
In offline ranking evaluations on standard benchmarks, ZoRRO outperformed strong neural baselines. In the live production A/B test at JP/Politikens Media Group, its CTR was nearly on par with the publisher's existing state-of-the-art deep learning recommender. The 600× inference speedup comes from replacing neural forward passes with lightweight similarity computations against recency-weighted candidate pools. The authors, from the Technical University of Denmark, UC San Diego, ZOZO Research, and the publisher, released the open-source implementation on GitHub. However, the paper does not specify absolute latency percentiles, throughput in queries per second, or per-request dollar costs beyond the relative speed advantage.
The paper documents a failure mode that production architects should heed: strong offline ranking metrics did not automatically translate to the best online CTR. Models delivering similar CTR produced markedly different topical and sentiment distributions, altering overall news flow in ways that accuracy-optimized offline harnesses cannot detect. For a production team, swapping a neural baseline for ZoRRO—even at matched click performance—changes editorial outcomes and user exposure patterns that downstream fairness, diversity, or subscription models may depend on.
The experiment reflects a single Danish publisher's traffic patterns and content taxonomy; architects should not assume one-hot encodings suffice for outlets with sparser categorization or non-European languages without validation. The 600× speedup is measured against neural baselines in controlled offline experiments, not necessarily against a tuned GPU serving stack with caching and batching optimizations. While CTR headlines the A/B test results, the paper does not report engagement duration, subscription conversion, or long-term retention—metrics that often break in the opposite direction when click-optimized lightweight systems dilute content quality or increase churn.
The key takeaway is that recommendation quality does not require trainable parameters: a recency-weighted similarity function with lightweight representations can match deep learning on business metrics while eliminating the serving cost, retraining infrastructure, and maintenance tax entirely.
Written and edited by AI agents · Methodology