GPT4Rec: Generative Recommender Framework

• GPT4Rec is a generative recommendation paradigm that transforms user purchase histories into natural language queries for BM25-based item retrieval.
• It employs beam search to generate multiple diverse queries, capturing the multifaceted nature of user interests and boosting metrics like Diversity@K and Coverage@K.
• Empirical evaluations show significant recall improvements on public datasets, while extensions address sequential, streaming, and multimodal recommendation challenges.

GPT4Rec refers to a family of generative recommender system frameworks that recast recommendation as a LLMing and retrieval problem. The foundational variant, as introduced by "GPT4Rec: A Generative Framework for Personalized Recommendation and User Interests Interpretation" (Li et al., 2023), utilizes generative LLMs to produce search queries based on user histories, followed by retrieval of items using search engine techniques. Subsequent variants and related works have extended the methodology to sequential recommendation, streaming graphs, multimodal input, and beyond-accuracy objectives, with continual innovation on query generation, retrieval, prompt tuning, and efficient adaptation.

1. Generative Recommendation Paradigm

GPT4Rec fundamentally shifts recommendation away from discriminative, ID-based modeling to a generative approach in the language domain. The primary workflow involves two stages: (1) a fine-tuned GPT-2 model generates hypothetical search queries from a user's past interacted item titles, and (2) a search engine, typically using BM25 lexical matching, retrieves items by searching these queries. Formally, GPT4Rec models the conditional distribution $P(q|W_u)$, where $W_u$ is a prompt constructed as "Previously, the customer has bought: <ITEM TITLES>. In the future, the customer wants to buy ...", generating natural language queries $q$ capturing semantic aspects of user interest.

This paradigm enables the inclusion of rich textual content of items into the user modeling process and unlocks the ability to generate recommendations for cold-start items having descriptive content but little historical interaction. The generated queries act as explicit, interpretable representations of user interests, which can be audited and manipulated for transparency.

2. Query Generation and Diversity via Beam Search

GPT4Rec deploys a multi-query generation mechanism—specifically, beam search—to reflect the multifaceted nature of user preferences:

• At each decoding step, a beam of $m$ candidate queries $Q^\ell$ is selected, expanding each candidate and retaining the top $m$ sequences by maximizing the generation score $S(W_u, q)$ for $|q| = \ell+1$.
• This approach permits the model to cover diverse user interests at multiple levels of granularity, increasing both the relevance and interpretability of recommendations.
• In practice, this multi-query framework provides strong empirical improvements in metrics that measure recommendation diversity and coverage, such as Diversity@K and Coverage@K, which utilize Jaccard similarity and categorical overlap between recommended and historical items.

Beam search in GPT4Rec ensures recommendations are not only accurate but also varied, aligning with goals for personalized, engaging recommendation lists.

3. Retrieval Mechanism and Adaptiveness

The second stage of GPT4Rec leverages the BM25 retrieval function parameterized by $k_1$ and $b$, selected via grid search:

• BM25 conducts lexical matching between the generated queries and available item titles.
• This retrieval strategy adapts robustly to growing or changing item inventories without need for retraining or embedding updates, as only textual content is required.
• The framework is naturally suited to address cold-start scenarios; newly introduced items with relevant semantic overlap are detectable and recommendable via generated queries.

A plausible implication is that the decoupling of query generation and item retrieval facilitates rapid system updates and integration with evolving product catalogs.

4. Performance Metrics and Numerical Results

Empirical evaluation of GPT4Rec on public datasets (Beauty and Electronics) demonstrates substantial advances over prior models:

Dataset	Method	Recall@K Improvement
Beauty	GPT4Rec-Beam	75.7%
Electronics	GPT4Rec-Beam	22.2%

• Diversity@K and Coverage@K are heightened using multi-query generation, showing a notable increase in recommendation breadth and representation of user multi-interests.
• Case studies reveal that GPT4Rec is capable of generating queries that are both specific and general, aligning precisely with users of narrow (e.g., one brand/product type) and broad tastes (multiple brands/types).

This strong numerical performance validates the framework’s ability to synthesize interpretable, diverse recommendations by leveraging content information.

5. Model Interpretability and User Interest Representation

The generative queries output by GPT4Rec are designed to be human-understandable and to faithfully summarize a user’s intent:

• Queries might explicitly express a need for items (e.g., "makeup palette") absent from prior interactions, indicating model capacity to infer latent interests.
• This property makes GPT4Rec recommendations particularly useful in domains where explanation or transparency is required, such as retail or content moderation.

A plausible implication is that this feature could facilitate regulatory compliance for explainable AI systems in sensitive application areas.

6. Extensions: Sequential, Streaming, and Multimodal Variants

Multiple subsequent works have evolved the GPT4Rec methodology:

• "Generative Sequential Recommendation with GPTRec" (Petrov et al., 2023) reformulates sequential recommendation as generative modeling, addressing large item vocabularies through SVD Tokenisation and introducing a Next‑K strategy for generating lists with inter-item dependencies.
• "GPT4Rec: Graph Prompt Tuning for Streaming Recommendation" (Zhang et al., 12 Jun 2024) adapts GPT4Rec for continual learning on streaming user–item graphs, employing disentangled views and node-, structure-, and view-level prompts for efficient dynamic adaptation.
• "Rec-GPT4V: Multimodal Recommendation with Large Vision-LLMs" (Liu et al., 13 Feb 2024) utilizes large vision-LLMs to bridge textual and image-based user histories, employing image summary generation to enable robust multimodal recommendations.
• "Aligning GPTRec with Beyond-Accuracy Goals with Reinforcement Learning" (Petrov et al., 7 Mar 2024) introduces a two-stage teacher-student plus RL fine-tuning protocol for optimizing diversity or popularity bias in the Next-K generative framework.

These extensions leverage the interpretability and adaptability of the original GPT4Rec design, applying it to increasingly complex recommendation scenarios.

7. Methodological Limitations and Future Directions

Several methodological constraints are noteworthy:

• Lexical retrieval via BM25 may limit semantic generalization, as demonstrated by later frameworks such as GLoSS (Acharya et al., 2 Jun 2025), which adopt dense semantic search and improved generative modeling.
• Scaling to very large item catalogs and real-time environments may require further algorithmic and engineering innovation, such as efficient retrieval indexing and parameter-efficient model adaptation (e.g., LoRA/QLoRA).
• GPT4Rec does not directly address multimodal or streaming dynamics beyond textual content in its foundational version, though streaming and multimodal extensions have recently emerged.

Potential future work may integrate semantic retrieval, multimodal fusion, and continual prompt adaptation to further increase robustness and relevance in real-world recommendation systems.

• "GPT4Rec: A Generative Framework for Personalized Recommendation and User Interests Interpretation" (Li et al., 2023)
• "Generative Sequential Recommendation with GPTRec" (Petrov et al., 2023)
• "GPT4Rec: Graph Prompt Tuning for Streaming Recommendation" (Zhang et al., 12 Jun 2024)
• "Rec-GPT4V: Multimodal Recommendation with Large Vision-LLMs" (Liu et al., 13 Feb 2024)
• "Aligning GPTRec with Beyond-Accuracy Goals with Reinforcement Learning" (Petrov et al., 7 Mar 2024)
• "GLoSS: Generative LLMs with Semantic Search for Sequential Recommendation" (Acharya et al., 2 Jun 2025)

GPT4Rec and its variants exemplify the generative-retrieval paradigm in recommender systems, offering interpretable, diverse, and adaptable recommendation pipelines, with ongoing research addressing limitations in retrieval semantics, scalability, and multi-modal reasoning.