Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5) - RecSys 2022

Forum topic · 小凯 · 2026-07-05

Summary

This forum post on zhichai.net introduces the RecSys 2022 paper 'Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5)'. P5 reformulates recommendation as a language processing task: a single pretrained language model is evaluated across diverse recommendation tasks via personalized natural-language prompts, unifying pretraining, prompting, and prediction in one framework. The post places the work in the Sequential Recommendation section of an academic reading list, alongside related entries such as BERT4Rec, EAGER, Mamba4Rec, and the SIGIR 2022 follow-up 'How to Index Item IDs for Recommendation Foundation Models'. Beyond the paper summary, the post provides broader context on how large language models are reshaping search, recommendation, and personalization pipelines, including discussions of retrieval-augmented generation, agentic search, evaluation practices, latency/cost constraints, and an engineering checklist for production deployment. The original paper is available via ACM DL at https://dl.acm.org/doi/abs/10.1145/3523227.3546767. Note that parts of the post use generic template text, so quantitative results should be verified against the original PDF.

Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5) — RecSys 2022

Overview

This entry from the zhichai.net academic reading list covers the RecSys 2022 paper "Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5)", filed under the *Sequential Recommendation* section.

Paper link: https://dl.acm.org/doi/abs/10.1145/3523227.3546767
Venue: RecSys 2022
Type: Academic paper

What P5 Proposes

P5 reframes recommendation as a language processing problem. Instead of building task-specific models, a single pretrained language model is adapted to handle diverse recommendation tasks (e.g., rating prediction, sequential recommendation, explanation generation) through personalized natural-language prompts. The paradigm unifies three stages:

1. Pretrain — a foundation model on user behavior data; 2. Personalized Prompt — task instructions conditioned on the user's history; 3. Predict — generating outputs directly as text.

This enables one model to serve multiple recommendation objectives with zero-shot or few-shot prompt specifications, laying groundwork for later "recommendation foundation models" and LLM-based generative recommendation (Gen-Rec).

Context from the Post

The post situates P5 within the broader shift toward LLM-era search and recommendation systems:

Classic cascades (recall → ranking → generation) are being rethought as LLMs blur the boundaries between retrieval, ranking, and generation;
Key trade-offs include inference budget, latency, hallucination risk, and evaluation credibility (offline nDCG vs. online satisfaction);
Agentic paradigms treat retrieval frequency and strategy themselves as learnable decisions.

The post also includes an engineering checklist (PII handling in training data, p99 latency budgets, interleaving experiments, retrieval poisoning defenses, per-query token cost) and a glossary of terms such as RAG, LTR, nDCG, and Agentic Search.

Related Entries Cross-Referenced

BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations
EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration
Efficient On-Device Session-Based Recommendation (ACM TORS)
How to Index Item IDs for Recommendation Foundation Models (P5 follow-up, SIGIR)
LLMCDSR: Enhancing Cross-Domain Sequential Recommendation with LLMs
Mamba4Rec: Efficient Sequential Recommendation with Selective State Space Models

Caveats

The post is partially template-generated: several sections (abstract quotes, methodology details) reuse boilerplate text, and the authors note that quantitative results should be verified against the original PDF. Specific benchmark numbers are not reproduced in the post itself.

Takeaways for Readers

1. Researchers: reproduce the core P5 comparisons and check for statistical significance and compute-cost reporting; 2. Engineers: evaluate P5-style prompt-based recommenders as pluggable modules against existing retrieval/ranking stacks; 3. Product managers: focus on user-perceivable benefits (latency, answer trustworthiness, multi-turn consistency) rather than offline metrics alone.