LLMCDSR: Enhancing Cross-Domain Sequential Recommendation with Large Language Models (ACM TOIS 2025)
Overview
| Field | Content | |---|---| | Title | LLMCDSR: Enhancing Cross-Domain Sequential Recommendation with Large Language Models | | Venue | ACM Transactions on Information Systems (TOIS), 2025 | | Source | https://dl.acm.org/doi/abs/10.1145/3715099 | | Category | Sequential Recommendation |
Key points
- LLMCDSR applies Large Language Models to cross-domain sequential recommendation, aiming to improve user intent understanding, efficiency, and scalability in large-scale personalized systems.
- The work responds to the limits of traditional pipelines that treat retrieval, ranking, and generation as disconnected stages, and to new expectations in the LLM era: natural language interaction, multi-hop reasoning, and up-to-date knowledge.
- It leverages the semantic priors and cold-start capabilities of LLMs, while acknowledging that online inference cost and hallucination risk require careful system design.
- BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations
- EAGER: Two-Stream Generative Recommender
- Efficient On-Device Session-Based Recommendation
- How to Index Item IDs for Recommendation Foundation Models (P5)
- Mamba4Rec: Efficient Sequential Recommendation with Selective State Space Models (arXiv:2403.03900)
- Original paper: LLMCDSR — ACM Digital Library
Context in the field
The post situates the paper within the evolution of recommendation and information retrieval:
1. IR side: from BM25 → BERT cross-encoders → dual-tower dense retrieval → late interaction → generative retrieval and LLM agents. Each generation balances an efficiency–effectiveness–maintainability triangle. 2. Recommendation side: from matrix factorization and deep CTR models → sequential Transformers → LLM instruction-following and generative recommendation (Gen-Rec). Core tensions: sparse user behavior, huge item catalogs, and multi-objective business trade-offs. 3. LLM-era shift: retrieval becomes iterative, verifiable, and plannable (RAG / agentic search); the new variables are inference budget and action space (whether to retrieve, how many steps, which tools to call).
Architecture discussion
The general technical pipeline described follows four stages:
1. Input and representation — encoding queries, documents, and user context into dense/sparse representations or structured prompts; 2. Core modules — retrievers, rerankers, planners, memory modules, and tool interfaces, composed in series or parallel; 3. Learning strategy — supervised fine-tuning, contrastive learning, distillation, reinforcement learning, and synthetic data bootstrapping; 4. Inference strategy — single-pass vs. iterative retrieval, parallel sub-queries, early stopping, and budget control.
Evaluation considerations
Typical benchmarks and metrics in this area include MS MARCO, BEIR, Natural Questions, and public recommendation datasets; metrics include nDCG@10, MRR, Recall@k, Hit@k, task success rate, latency, and token cost. Baselines usually cover BM25, dense retrieval, cross-encoder reranking, and retrieval-free LLMs.
> Note: The exact quantitative results of LLMCDSR should be verified against the original PDF. This post is based on the abstract and public metadata.
Takeaways for search/rec practitioners
1. Architecture: cascaded retrieve–rerank–generate remains mainstream, but agentic paradigms make retrieval strategy itself learnable. 2. Data: high-quality instruction data and click/session logs matter equally; synthetic data must guard against leakage and distribution shift. 3. Evaluation: the gap between offline metrics and online satisfaction is widening; LLM-as-judge needs cross-validation with human assessment. 4. Production: latency, cost, explainability, and safety are hard constraints — optimizing academic benchmarks alone is insufficient.