Research map / Large language models / Language, safety and interpretability
Principled Synthetic Data Enables the First Scaling Laws for LLMs in Recommendation
Benyu Zhang, Qiang Zhang, Jianpeng Cheng, Hong-You Chen, Qifei wang, Wei Sun, Shen Li, Jia Li, Jiahao Wu, Xiangjun Fan, Hong Yan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train sequential recommenders on synthetic UIH instead of real logs (TSTR vs TRTR) | GRU4Rec, NARM, STAMP, SASRec · Merrec test, Recall@10/100/1000 | Recall@K: +130% on recall@100 for SasRec | Improved |
| add collaborative filtering data alongside UIH data (ablation) | Qwen3-4B CPT · UIH Full (avg) | asymptotic loss L_inf: 0.95 -> 0.66 | Improved |
| train on UIH data alone and evaluate on CF (ablation) | Qwen3-4B CPT · CF Both Unseen | scaling exponent alpha: alpha ≈ 0 | No clear effect |
| include CF data in the mixture (ablation) | Qwen3-4B CPT · CF Both Unseen | asymptotic loss L_inf: 1.30 -> 0.35 | Improved |
| include item-text data in the mixture (ablation) | Qwen3-4B CPT · Item-Text | asymptotic loss L_inf: ≈3 -> ≈1.2 | Improved |
| include item-text data in the mixture (ablation) | Qwen3-4B CPT · UIH Full (avg) | asymptotic loss L_inf: 0.66 -> 0.76 | Hurt |
| raise UIH mixture ratio so reduced UIH data repeats about 16 times or more (ablation) | Qwen3-4B CPT · held-out UIH perplexity | perplexity | Hurt |
| scale model size from 0.6B to 8B with fixed 163B tokens (ablation) | Qwen3 CPT · UIH OOD | perplexity: 2.01 -> 2.10 | Hurt |
| scale model size from 0.6B to 4B and beyond (ablation) | Qwen3 CPT · Collaborative filtering | perplexity: saturates at ~1.47 PPL at 4B | No clear effect |
| scale model size from 0.6B to 8B (ablation) | Qwen3 CPT · General domain | scaling exponent alpha: alpha 0.02-0.03 | No clear effect |
| scale model size from 0.6B to 8B (ablation) | Qwen3 CPT · UIH | scaling exponent alpha: 0.45 -> 0.59 | Improved |
| apply method to Gemma 3 family on proprietary social media dataset | Layered synthetic CPT · Internal dataset | scaling-law fit RMSE: RMSE < 0.035 | Improved |
Closest papers by meaning
- Can LLMs Outshine Conventional Recommenders? A Comparative Evaluation (NeurIPS 2025 · similarity 0.90)
- LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.90)
- SLMRec: Distilling Large Language Models into Small for Sequential Recommendation (ICLR 2025 · similarity 0.90)
- LLM4RSR: Large Language Models as Data Correctors for Robust Sequential Recommendation (AAAI 2025 · similarity 0.89)
- Wukong: Towards a Scaling Law for Large-Scale Recommendation (ICML 2024 · similarity 0.88)
- Language Ranker: A Lightweight Ranking framework for LLM Decoding (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
