Recommendation as Instruction Following: An LLM-Empowered Recommendation Approach (ACM, Dec 2024)

• Source: https://dl.acm.org/doi/abs/10.1145/3708882
• Published: December 2024, ACM
• Type: Academic paper

Overview

This paper proposes reframing the recommendation task as instruction following with large language models. Rather than treating recommendation purely as a ranking or retrieval problem over item catalogs, the approach encodes user intent, context, and recommendation goals as natural-language instructions that an LLM can interpret and act upon.

Context

Traditional recommender pipelines (candidate generation → ranking → re-ranking) have long contended with:

• Sparse user behavior signals and extremely large item catalogs
• Weak semantic understanding and poor cold-start performance
• Rigid separation between retrieval, ranking, and generation stages

LLMs introduce semantic priors and reasoning capabilities that can address cold-start and intent understanding, but they also bring new constraints: online inference cost, latency budgets, and hallucination risk.

Key Points

• Problem framing: Recommendation is treated as an instruction-following task, aligning it with the broader paradigm of LLM instruction tuning.
• Positioning: The work belongs to the generative recommendation (Gen-Rec) line, evolving from matrix factorization → deep CTR models → sequential Transformers → LLM-based recommenders.
• System implications: LLM-era systems treat inference budget and action space (whether/when to retrieve, which tools to call) as first-class design variables.
• Engineering constraints: Latency (p99), token cost, caching of hot queries, embedding version management, and output safety filtering are hard requirements for production deployment.

Reader Guidance

• Quantitative results, datasets, and baselines should be verified against the original PDF via the ACM Digital Library; this entry is based on public metadata and abstract information.
• Related threads: LLM-based representation learning for recommendation, generative retrieval, agentic search, and instruction-tuned rankers.

Takeaways for Practitioners

1. Researchers: Check whether evaluations report statistical significance and compute cost alongside accuracy metrics. 2. Engineers: Identify pluggable components (encoders, re-rankers, planners) and assess integration cost with existing retrieval/ranking stacks. 3. Product teams: Focus on user-perceivable benefits—latency, answer trustworthiness, multi-turn consistency—rather than offline metrics alone.