Recommendation as Instruction Following: An LLM-Empowered Recommendation Approach (arXiv, May 2023)
Paper: https://arxiv.org/abs/2305.07001
Authors: Junjie Zhang, Ruobing Xie, Yupeng Hou, Wayne Xin Zhao, Leyu Lin, Ji-Rong Wen (Renmin University of China / Kuaishou collaboration)
Venue: arXiv preprint, May 2023
Key points
- Reframes recommendation as an instruction-following problem: user interactions, item metadata, and recommendation tasks are verbalized into natural language instructions that a large language model can directly consume and act upon.
- Proposes a single LLM-based framework that can be tuned on instruction-formatted recommendation data, unifying multiple recommendation tasks (e.g., rating prediction, sequential/item recommendation) under one model instead of task-specific pipelines.
- Emphasizes zero-shot generalization: after instruction tuning, the model can handle unseen recommendation tasks and transfer across domains, mitigating cold-start and data-sparsity issues common in collaborative filtering.
- Leverages the semantic prior of pretrained LLMs (world knowledge, language understanding of item text) to complement or replace traditional ID-based embedding approaches.
- An early, influential entry in the LLM-for-recommendation line of work, alongside later efforts such as decoder-only foundation models for personalized ranking and language-item retrieval bridging.
- Architecture: instruction-tuned LLMs can act as unified recommenders, but latency and inference cost remain hard constraints for production deployment.
- Data: the quality of instruction-formatted recommendation data is as important as model scale; watch for distribution shift between training tasks and target scenarios.
- Evaluation: text-generation metrics must be paired with ranking metrics (Recall@k, nDCG) and online signals.
- Open problems: hallucinated or non-existent item outputs, catalog alignment, and multilingual/multimodal extension.
- 360Brew: A Decoder-only Foundation Model for Personalized Ranking
- Bridging Language and Items for Retrieval and Recommendation (Mar 2024)
- Data-efficient Fine-tuning for LLM-based Recommendation, SIGIR 2024
Motivation
Traditional recommenders (matrix factorization, deep CTR models, sequential Transformers) learn from user–item interaction IDs. They struggle with sparse feedback, cold-start items and users, and require separate models per task. Meanwhile, LLMs exhibit strong generalization from instruction tuning in NLP. This paper asks: can recommendation itself be cast as instruction following, so that one LLM serves many recommendation scenarios?
Approach outline
1. Task verbalization — encode user history, candidate items, and the target task as natural language instructions/templates. 2. Instruction tuning — fine-tune an LLM on instruction–response pairs built from recommendation datasets so the model learns to produce recommendation outputs (e.g., predicted ratings or next items) as text. 3. Inference — at recommendation time, format the user's context as an instruction and let the tuned LLM generate the answer, enabling zero-shot handling of new tasks and domains.
Evaluation
The paper evaluates on public recommendation benchmarks, comparing against traditional collaborative filtering and sequential baselines, and specifically tests zero-shot/transfer settings where the model faces tasks or domains not seen during tuning. Readers should consult the original PDF tables for exact quantitative results.