Principled Synthetic Data Enables the First Scaling Laws for LLMs in Recommendation

Large Language Models (LLMs) represent a promising frontier for recommender systems, yet their development has been impeded by the absence of predictable scaling laws, which are crucial for guiding research and optimizing resource allocation. We hypothesize that this may be attributed to the inherent noise, bias, and incompleteness of raw user interaction data in prior continual pre-training (CPT) efforts. This paper introduces a novel, layered framework for generating high-quality synthetic data that circumvents such issues by creating a curated, pedagogical curriculum for the LLM. We provide powerful, direct evidence for the utility of our curriculum by showing that standard sequential models trained on our principled synthetic data significantly outperform ($+130\%$ on recall@100 for SasRec) models trained on real data in downstream ranking tasks, demonstrating its superiority for learning generalizable user preference patterns. Building on this, we empirically demonstrate, for the first time, robust power-law scaling for an LLM that is continually pre-trained on our high-quality, recommendation-specific data. Our experiments reveal consistent and predictable perplexity reduction across multiple synthetic data modalities. These findings establish a foundational methodology for reliable scaling LLM capabilities in the recommendation domain, thereby shifting the research focus from mitigating data deficiencies to leveraging high-quality, structured information.

Recommendation systems — the technology behind "you might also like" suggestions — are increasingly powered by large language models (LLMs), the same AI behind modern chatbots. But unlike in language tasks, researchers have been unable to predict how these models improve as they are trained on more data, making development expensive and unreliable. We discovered that the problem lies not in the models themselves, but in the data they learn from. Real user activity logs are full of hidden biases: for example, people tend to click on whatever appears first in a list, making it look like top-ranked items are always the best. Models trained on this data learn these flaws rather than genuine user preferences. To fix this, we built a system that creates clean, structured training data from scratch — like designing a well-organized textbook instead of handing students a pile of messy notes. Models trained on our synthetic data outperformed those trained on real data, and for the first time, we observed the kind of predictable, reliable scaling laws that have guided progress in other areas of AI. This means researchers can now forecast how much data and computing power they need to reach a target quality level, replacing guesswork with science.