In natural language processing, scaling laws have provided a predictable playbook: increase compute, dataset size, and parameter count, and downstream loss drops according to a clean power law. However, deep learning recommendation systems (RecSys) have historically failed to exhibit these predictable gains. Instead of scaling dense compute layers, traditional recommendation models scale by expanding sparse embedding tables to accommodate billions of user and item IDs. This memory-bound paradigm results in diminishing returns and offers no clear formula for forecasting performance.
To bridge this gap, a new wave of research focuses on establishing predictable scaling laws for recommendation systems. By transitioning from ID-centric architectures to dense feature interaction stacks and generative foundation models, researchers are unlocking power-law performance improvements that mirror those seen in large language models (LLMs).
Why Traditional Recommendation Fails to Scale
To understand why recommendation scaling laws are difficult to establish, we must look at the constraints unique to the domain: data noise, architectural mismatches, and strict inference latency limits.
Data Noise and Sparsity
Real-world user interaction logs are incredibly noisy and sparse. As Benyu Zhang et al. demonstrate in Principled Synthetic Data Enables the First Scaling Laws for LLMs in Recommendation, this inherent bias breaks predictable power laws. When models are trained on raw, unpurified user logs, the scaling signal is drowned out by behavioral noise. By utilizing a layered synthetic data framework to "purify" user signals, the authors successfully unlocked the first robust power-law scaling trajectories for recommendation models, showing that models trained on clean synthetic data can significantly outperform those trained on real logs.
Architecture Mismatch
Standard recommendation architectures are not optimized to leverage increased model capacity. To address this, Buyun Zhang et al. introduced Wukong in Wukong: Towards a Scaling Law for Large-Scale Recommendation, showing that scaling dense interaction layers (via stacked Factorization Machines) rather than merely expanding embedding table memory yields highly predictable scaling across multiple orders of magnitude. Similarly, Wei Guo et al. developed FuXi-α: Scaling Recommendation Model with Feature Interaction Enhanced Transformer, which uses a multi-channel attention mechanism to ensure performance scales predictably as model depth and hidden dimensions grow.
The Latency Constraint
Unlike LLMs where parameter count dominates resource discussions, recommendation systems operate under strict millisecond-level online inference constraints. As noted by Xiao Lv et al. in MARM: Unlocking the Recommendation Cache Scaling-Law through Memory Augmentation and Scalable Complexity, computational cost (FLOPs) and inference latency are the primary bottlenecks rather than raw parameter count. The authors propose a cache-based scaling law that allows multi-layer sequence modeling to scale with near-linear complexity, making large-scale recommendation models practical for production environments.
The Shift to Generative and Foundation Recommendation
The most dramatic breakthrough in establishing recommendation scaling laws comes from treating recommendation as a sequential transduction task, similar to generative language modeling.
In a pioneering single-author work, Rui Li introduces Generative Recommenders (GRs) in Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations. By replacing traditional architectures with a highly efficient sequential transducer (HSTU), Rui Li scaled recommendation models to 1.5 trillion parameters, proving that recommendation quality follows a strict power-law scaling relationship with compute across several orders of magnitude.
This transition toward Generative Large Recommendation Models (GLRMs) is further formalized by Hao Wang et al. in Generative Large Recommendation Models: Emerging Trends in LLMs for Recommendation, where the authors show that data compression entropy can serve as a reliable predictor for ranking and retrieval performance as model capacity scales.
Traditional RecSys: [User/Item IDs] ──> [Massive Sparse Embedding Tables] ──> Diminishing Returns
Generative RecSys: [User Actions] ──> [Dense Sequential Transducers] ──> Power-Law Scaling
Furthermore, scaling up these models yields massive benefits for data scarcity, sample efficiency, and transferability:
- Sample Efficiency: Jianghao Lin et al. demonstrate in Large Language Models Make Sample-Efficient Recommender Systems that as the underlying language model size scales up, its sample efficiency improves dramatically. Larger models leverage their broader internal knowledge base to map sparse user histories to potential future interests, allowing a model trained on only 10% of the data to outperform traditional baselines.
- Cold-Start and Long-Tail Performance: In Scaling Law of Large Sequential Recommendation Models, Gaowei Zhang et al. show that the performance of massive sequential recommendation models can be accurately forecasted from smaller runs. Crucially, scaling up consistently mitigates cold-start and long-tail item prediction challenges, proving that model capacity can compensate for data scarcity.
- Cross-Platform Transferability: Moving away from rigid ID features, Jie Wang et al. introduce TransRec: Learning Transferable Recommendation from Mixture-of-Modality Feedback, proving that foundation recommendation models scale consistently with the volume of pre-training interaction data across text and images, capturing universal behavioral patterns applicable across diverse platforms.
- Upper Limits of Text-Based Encoders: Pushing the boundaries of text-based collaborative filtering, Ruyu Li et al. show in Exploring the Upper Limits of Text-Based Collaborative Filtering Using Large Language Models: Discoveries and Insights that recommendation performance continues to scale up to at least 175 billion parameters, though task-specific fine-tuning remains essential to unlock full performance.
Parallel Frontiers: Scaling Laws in Reinforcement Learning
Because recommendation systems are inherently sequential feedback loops, their scaling challenges closely mirror those found in Reinforcement Learning (RL). Historically, RL has also lacked systematic scaling laws due to the complexities of self-play, value optimization, and environment variance. Recent work has begun mapping these trajectories, offering valuable cross-domain insights for scaling sequential decision models.
Oren Neumann et al. explore this in Size Scaling in Self-Play Reinforcement Learning and Scaling Laws for a Multi-Agent Reinforcement Learning Model, demonstrating that self-play agents (like AlphaZero) do follow predictable power laws relative to model size and compute budgets once properly isolated, allowing researchers to derive mathematically optimal model sizes for a given compute budget.
To make scaling compute-optimal in value-based settings, Preston Fu et al. introduce a formal framework in Compute-Optimal Scaling for Value-Based Deep RL. The authors identify a key phenomenon called TD-overfitting, where small models collapse under large batch sizes, providing a roadmap to balance model capacity and update frequency.
When scaling RL for LLM post-training, Fnu Devvrit et al. present ScaleRL in The Art of Scaling Reinforcement Learning Compute for LLMs, utilizing sigmoidal curves to accurately forecast large-scale training success from smaller initial runs. Finally, Zhoujun Cheng et al. provide the IsoCompute Playbook: Optimally Scaling Sampling Compute for LLM RL, demonstrating that the compute-optimal strategy requires predictably scaling parallel rollouts per problem to mitigate task interference.
Together, these advancements in both recommendation architectures and reinforcement learning frameworks demonstrate that sequential, feedback-driven systems are not inherently immune to scaling laws. By shifting toward dense interaction layers, generative architectures, and rigorous compute-optimal resource allocation, researchers can now forecast and achieve predictable performance gains at scale.
Key Foundational Work
Exploring the Upper Limits of Text-Based Collaborative Filtering Using Large Language Models: Discoveries and Insights (2023) — contributed 1 concept elements to this direction.
Size Scaling in Self-Play Reinforcement Learning (2022) — contributed 1 concept elements to this direction.
Scaling Laws for a Multi-Agent Reinforcement Learning Model (2023) — contributed 1 concept elements to this direction.