CONCEPT Cited by 1 source
Scaling laws for recommenders¶
Definition¶
Scaling laws for recommenders refers to the empirical observation that generative recommendation models exhibit power-law loss reduction with increasing model capacity — mirroring the scaling laws established for large language models (Kaplan et al. 2020, Chinchilla 2022).
Key Findings (Netflix GenPage)¶
Netflix's GenPage swept model sizes from 120M to 900M parameters and found:
- Both pretraining loss and WBC post-training loss decrease in a power-law-like fashion with model size
- However, context enrichment dominates capacity scaling in their regime: enriching the user prompt reduced WBC loss by ~6.9%, while a ~7.5× model size increase yielded only ~1.3% reduction
- The implication: personalization quality is bottlenecked by available information (what the model sees) before it's bottlenecked by capacity (how big the model is)
This suggests that for industry-scale recommender systems, investing in richer input signals (more history, more metadata, better tokenization) provides stronger returns than simply scaling parameters — at least until the context is saturated (Source: sources/2026-06-29-netflix-genpage-generative-homepage-construction).