Deep Recommender Research

DuoKD introduces a dual positive-negative knowledge extraction strategy based on LLMs, integrated with a knowledge distillation mechanism for efficient models.

Memory Scaling Innovations

• 🔥 MSA: Memory Sparse Attention: MSA presents an end-to-end trainable, scalable sparse latent-state memory framework...

Key breakthrough for long-horizon sequential recommendation:

• Attention bottleneck limits LLMs to 128K–1M contexts despite long-term memory needs.
  -...

This AI system practically deploys a Skill Gap Severity Index to identify missing and underserved skills for target roles, paired with a Customized Learning Plan Generator for tailored recommendations.

Discover GraphER, a method splitting retrieval for efficiency in recsys:

• Offline indexing: Independently enriches data objects
• Query time:...

Addressing limitations of existing methods, researchers propose DNSC, a Diffusion-enhanced Negative Sampling model in multimodal contrastive learning for recommendation.

Retrieval and Ranking Innovations

• 🔥 ColdRAG: During inference, ColdRAG performs query-aware multi-hop reasoning over KG to adaptively retrieve...

• Six AI chips released in two years for blazing-fast MTIA architecture
• Meta's AI split into four quadrants: training vs. inference, crossed with recommendation models vs. GenAI
• Signals rapid hardware push for low-latency RecSys production infra

Emerging generative rec models push real-world deployment:

• On-device LLMs for sequential rec introduce tokenization normalization to fix transformer...

One-layer transformers with simplified position-only attention, trained via gradient descent, provably recover all teacher models from a specific class. This provides key theoretical foundations for model-efficient sequential architectures.

Multimodal RecSys Architectures

• 🔥 E-commerce Visual Recommendation: Employs a Transformer to model user behavior sequences and CNNs to extract...

New arXiv paper [2603.21541] derives sharper generalization error bounds for Transformer models using offset Rademacher complexity. Vital theoretical progress for robust deep recsys training.

LLM-guided data distillation targets explainable recommender systems, referencing PETER's transformer-based architecture with attention masking.

Key multimodal recsys advances:

• Transformer models user behavior sequences
• CNNs extract visual features from product images
• Attention mechanism enables personalized e-commerce visual search

New study leverages Cross-Encoder BERT to compute semantic relevance between users and jobs, enabling explainable and scalable job recommendation systems—building on prior findings.

Generative Recommendation Insights

• 🔥 How Well Does Generative Recommendation Generalize?: New paper examines generalization of generative...

Mamba-3, an open-source State Space Model, beats Transformers. At 1.5B parameters, Mamba-3 SISO delivers the fastest prefill + decode latency across all sequence lengths, topping Mamba-2 and Gated DeltaNet.

Emerging trend in advanced graph methods for multimodal recommendation:

• Conditional Graph Diffusion with granular-ball representation targets MRS...

New paper proposes a subgoal-driven framework to improve long-horizon LLM agents. Essential reading for advancing reliability in extended tasks.

New paper asks: How Well Does Generative Recommendation Generalize? – probing robustness across domains and setups. Join the discussion on this key research page.