Research map / Architectures and efficiency / Attention and transformers
Massive Memorization with Hundreds of Trillions of Parameters for Sequential Transducer Generative Recommenders
Zhimin Chen, Chenyu Zhao, Ka Chun Mo, Yunjiang Jiang, Jane H. Lee, Khushhall Chandra Mahajan, Ning Jiang, Kai Ren, Charlie Li, Wen-Yun Yang
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 |
|---|---|---|---|
| Add generative reconstruction loss (ablation) | VISTA · Industrial production eval NE | Eval NE: C-Task -0.29% -> -0.40%; E1-Task -1.29% vs -1.19% | Mixed |
| Add generative reconstruction loss (ablation) | VISTA · Amazon-Electronics | AUC/NE: +0.22% AUC, 1.11% NE | Improved |
| Use 256-dim instead of 128-dim embeddings (ablation) | VISTA · Industrial eval NE C-Task | Eval NE: -0.29% -> -0.40% | Improved |
| Use 128 seeds instead of 64 (ablation) | VISTA · Industrial eval NE | Eval NE: C-Task -0.37% -> -0.40%; E2 -3.01% vs -2.98% | Mixed |
| Increase number of seed embeddings (ablation) | VISTA · Minimal Production | AUC/NE | Improved |
| Use quasi-linear attention instead of softmax attention (ablation) | VISTA · Minimal Production training time | epoch time | Improved |
| Use quasi-linear attention (longer seq, more layers) (ablation) | VISTA · Industrial data | Eval NE: -0.13% eval NE, +5% QPS | Improved |
| Use VISTA instead of HSTU baselines | HSTU · Amazon-Electronics | AUC: 0.884 -> 0.886 | Improved |
| Use VISTA instead of HSTU baselines | HSTU · Minimal Production | AUC: 0.668 vs 0.627 | No clear effect |
| Use VISTA instead of HSTU | HSTU · KuaiRand-1K | NE: 0.863 -> 0.863 | No clear effect |
| Add VISTA to HSTU production model | HSTU · Online A/B 15 days | C-Task: +0.5%, 94% GPU reduction | Improved |
| Causal vs full self attention for user history (ablation) | VISTA · summarization | quality | No clear effect |
Closest papers by meaning
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024 · similarity 0.90)
- LIME: Link-based User-item Interaction Modeling with Decoupled XOR Attention for Efficient Test Time Scaling (NeurIPS 2026 · similarity 0.89)
- LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.88)
- SLMRec: Distilling Large Language Models into Small for Sequential Recommendation (ICLR 2025 · similarity 0.88)
- MARS: Multi-resolution Adaptive Routing for Sequential Recommendation (NeurIPS 2026 · similarity 0.88)
- Long-Sequence Recommendation Models Need Decoupled Embeddings (ICLR 2025 · similarity 0.87)
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.
