Source-linked AI summary
Large Language Models for Generative Recommendation: A Survey and Visionary Discussions
Lei Li, Yongfeng Zhang, Dugang Liu, Li Chen
TL;DR
Conventional LLM-based recommendation often uses LLMs as components of multi-stage pipelines, potentially underusing their generative power. This survey defines generative recommendation, reviews its methods and progress, and discusses implementation and future directions, while identifying practical constraints such as ID representation and evaluation limitations.
Problem
Existing work often treats LLMs as components of conventional recommendation pipelines rather than directly leveraging their generative power for recommendation.
Method
The survey defines generative recommendation, develops an LLM-based recommendation taxonomy, and reviews formulations, ID creation methods, tasks, evaluation, and open challenges.
Results
The survey systematically summarizes LLM-based generative recommendation and provides general formulations, implementation guidance, advantages over traditional recommender systems, and future research directions.
Takeaways & Limitations
Generative recommendation offers a single-stage approach in which an LLM directly generates recommended items from the complete item pool.
Takeaways & Limitations
Long user or item IDs can make generation computationally expensive and complicate exact database matching, potentially requiring comparisons with individual items.
Abstract
from arXiv · showhide
Large language models (LLM) not only have revolutionized the field of natural language processing (NLP) but also have the potential to reshape many other fields, e.g., recommender systems (RS). However, most of the related work treats an LLM as a component of the conventional recommendation pipeline (e.g., as a feature extractor), which may not be able to fully leverage the generative power of LLM. Instead of separating the recommendation process into multiple stages, such as score computation and re-ranking, this process can be simplified to one stage with LLM: directly generating recommendations from the complete pool of items. This survey reviews the progress, methods, and future directions of LLM-based generative recommendation by examining three questions: 1) What generative recommendation is, 2) Why RS should advance to generative recommendation, and 3) How to implement LLM-based generative recommendation for various RS tasks. We hope that this survey can provide the context and guidance needed to explore this interesting and emerging topic.
1. Introduction
This survey frames LLM-based generative recommendation as a shift from conventional multi-stage recommendation toward direct generation, and organizes the topic around definitions, methods, tasks, and future directions.
- ID representation: The survey defines IDs broadly as token sequences that uniquely identify users or items, including numerical tokens, word tokens, and item metadata.This formulation makes IDs compatible with natural-language environments and LLMs.
- Motivation: Traditional recommender systems reduce millions of items to a few hundred candidates before applying advanced recommendation algorithms.This multi-stage filtering paradigm addresses the computational burden of scoring every item.
- Definition: Generative recommendation directly generates recommendations or recommendation-related content without calculating each candidate’s ranking score one by one.The survey presents this as a single-stage alternative enabled by LLM generative power.
- Survey scope: The survey focuses specifically on generative recommendation and develops a recommendation-oriented taxonomy rather than adopting an NLP-derived LLM taxonomy.It distinguishes this organization from recent surveys on LLM-based recommendation.
- Contributions: The survey contributes a systematic review, task formulations, generalized ID definitions, and promising directions for LLM-based generative recommendation.It positions these contributions as guidance for future research in an under-explored area.
2. Why Generative Recommendation
Generative recommendation is motivated by the computational cost of multi-stage filtering and uses LLMs to generate recommendations directly from the item pool. Finite token sequences and bounded next-token probabilities make this direct generation computationally feasible.
- Computational motivation: Scoring every item is computationally expensive when recommendation platforms contain prohibitively many items, motivating multi-stage candidate narrowing.Traditional systems use simple models or filtering strategies early in the pipeline before applying more advanced methods.
- Single-stage generation: Generative recommendation advances the typical multiple-stage pipeline to one stage by directly generating items for recommendation with an LLM.The survey contrasts this pipeline with traditional recommendation in Figure 1.
- ID capacity: Using 1000 tokens for 10-token IDs yields 1000^10 = 10^30 unique item IDs, making a large item space representable with finite tokens.The passage describes numerical, word, and out-of-vocabulary tokens as possible ID components.
- Generation feasibility: Beam search keeps each generation step bounded by the 1000-token vocabulary, making direct recommendation generation computationally possible.The bounded probability vector supports generation from the item pool rather than explicit per-item scoring.
3. ID Creation Methods
LLM-compatible recommendation IDs must be concise and unique, motivating token sequences that encode collaborative or semantic structure while distinguishing users and items.
- ID design requirements: Long metadata-based IDs are costly to generate and can be ambiguous, making exact database matching difficult.Identical names may refer to unrelated items, while related items may have different titles.
- ID design requirements: Embedding IDs can distinguish entities but require substantial memory and are out-of-vocabulary tokens for LLMs.
- Token-sequence IDs: Token-sequence IDs use a small vocabulary to represent many entities, sharing tokens among similar users or items while preserving uniqueness.This design aims to fit recommendation IDs into the natural-language environment of LLMs.
- ID creation approaches: SVD-based IDs transform item embeddings from interaction data through normalization, noise addition, quantization, and offset adjustment into integer sequences.Noise addition helps ensure that item IDs are not identical.
- ID creation approaches: Hierarchical-tree IDs recursively cluster items using co-occurrence graphs so that tree nodes compose item identifiers.
- ID creation approaches: RQ-VAE IDs repeatedly quantize item embeddings with codebooks, using each selected codeword as a token in the item ID sequence.Residual vectors are updated after each nearest-codebook lookup.
- ID creation approaches: Other reported ID strategies include sequential indexing from interaction histories and semantic indexing from item metadata.
4. How to Do Generative Recommendation
LLM-based generative recommendation uses task prompts containing user and item information, but task design must account for context limits and distinguish generation from per-item scoring.
- Task formulation: Prompt templates describe the recommendation task and insert user or item IDs so one LLM can perform different generative recommendation tasks.Relevant tasks and models are summarized in Table 2.
- Rating prediction: Rating prediction prompts an LLM with a user and item ID to estimate how the user would rate that item.
- Rating prediction: Explicit ratings can be impractical because users may not rate every interacted item, motivating recommendation from implicit feedback.
- Top-N recommendation: Top-N recommendation selects N items for a user, whereas conventional systems compute a score for each candidate before choosing the highest-ranked items.
- Top-N recommendation: LLM context limits prevent supplying all items, leading to straightforward recommendation from user information or selective recommendation using a restricted candidate set.
- Top-N recommendation: Beam search can generate several item IDs, producing a recommendation list of N items.
- Generative versus discriminative recommendation: Generating yes-or-no answers for each item remains discriminative recommendation because it still produces an answer or score per item.
4.3. Sequential Recommendation
Sequential recommendation predicts a user’s next interacted item from chronologically ordered history, commonly prompting an LLM with recent item IDs.
- Task definition: Sequential recommendation predicts the next item a user will interact with from chronologically ordered past interactions.
- LLM formulation: A sequential prompt lists item IDs in interaction order and asks the LLM to generate the next item ID.An example output is an item ID such as “6789”.
- LLM formulation: Older interactions can be truncated from the prompt to reduce inference time.
- Research landscape: The task has attracted numerous LLM-based recommendation models, while some systems use LLMs to generate candidates for later filtering.
4.4. Explainable Recommendation
Explainable recommendation uses LLMs to generate natural-language reasons for recommending an item, optionally guided by item features and varied prompting strategies.
- Explanation formats: Recommendation explanations can use explicit item features or visual highlights to communicate why an item was recommended.
- Natural-language explanation generation: Natural-language explanation generation asks an LLM to produce a sentence or paragraph explaining why an item is recommended to a user.
- Prompt construction: A prompt containing user and item IDs can request an explanation, while adding an item feature focuses the explanation on a specified aspect.User reviews can provide ground-truth explanations.
- Prompting strategies: Researchers have used keyword triggers, chain-of-thought prompting, and continuous prompt vectors in explanation generation.
4.5. Review Generation
Review generation uses prompts containing a user and interacted item so an LLM can generate a review, potentially supporting downstream recommendation research. The survey also identifies review summarization settings that may be more useful than summarizing a review for its author.
- An LLM can generate a product or service review from a prompt containing a user and an interacted item.The formulation is illustrated with a prompt such as “generate a review for user_1234 about item_5678.”
- Generated reviews could make it easier for users to leave comments after purchases, viewing experiences, or rides.The resulting data may support explainable and conversational recommendation research.
- Review summarization can help users quickly understand an item’s pros and cons when reading a long review is costly.
- Summarizing a review for another user or aggregating multiple users’ opinions may be more useful than summarizing it for the original reviewer.
4.7. Conversational Recommendation
Conversational recommendation uses multiple dialogue rounds to incorporate users’ natural-language preferences and feedback. The survey describes dialogue-labeling, tool-calling, and direct-chat approaches, alongside offline and human-oriented evaluation considerations.
- Conversational recommendation recommends items through multiple rounds in which users can state preferences or reject recommendations in natural language.
- Some approaches mark dialogue speakers with “USER” and “SYSTEM” labels before providing the session to an LLM.
- Other approaches instruct the LLM to call traditional recommenders or SQL tools to narrow candidate items.
- Conversational recommendation can be evaluated with ranking metrics such as NDCG, precision, and recall, while online A/B tests reflect actual user interactions.
- For generated conversational content, BLEU and ROUGE measure overlap with ground-truth text, but text matching does not capture every quality dimension.The survey notes that more advanced standard metrics and human evaluation are needed.
5. Challenges and Opportunities
The survey identifies challenges including simulator validity, hallucination, bias, explainability, controllability, and efficiency, while highlighting opportunities for tool-using agents, multimodal systems, and cold-start recommendation. These directions require reliable grounding, fairness, transparency, controllable outputs, and efficient inference.
- 5.1. LLM-based Agents: Simulator data are useful only when they align with users’ true preferences, while perfect simulation would make recommendation algorithms unnecessary.
- 5.1. LLM-based Agents: LLM-based agents could use tools, APIs, and expert models to build and dynamically revise personalized itineraries with real-time information.The survey gives trip planning, routing, and nearby-service recommendation as examples.
- 5.2. Hallucination: Hallucinated recommendations may refer to nonexistent items, so item indexing and retrieval augmentation are proposed to ground outputs in the item database.The survey also notes that indexing and retrieval can be integrated.
- 5.3. Bias and Fairness: LLM-based recommendation faces content and recommendation bias, including gendered explanation lengths, popularity effects, and demographic differences in recommendations.The boundary between bias and personalization remains an open question.
- 5.4. Transparency and Explainability: Internal explainability for LLM-based recommendations remains largely unexplored, and knowledge-graph alignment is described as preliminary.
- 5.5. Controllability: Controllability is needed for both explanation features and recommended-item attributes such as price, color, brand, and price range.
- 5.6. Inference Efficiency: Because LLMs are parameter-heavy and recommendation is latency-sensitive, inference efficiency is especially important.The survey contrasts inference efficiency with offline training efficiency and mentions caching and parameter-efficient tuning approaches.
- 5.7. Multimodal Recommendation: Tokenized images, videos, and audio can extend LLM-based recommendation to multimodal tasks and even support generating new items when repositories lack suitable options.
6. Conclusions
The survey reviews LLM-based generative recommendation, formulates its tasks, explains its advantages over traditional recommender systems, generalizes item IDs, and identifies future research directions. It positions LLM–RS integration as a route toward personalized services across scenarios.
- The survey reviews recent progress and provides general formulations for LLM-based generative recommendation tasks.
- It explains advantages over traditional recommender systems, generalizes item IDs, and summarizes ID creation methods.
- The survey identifies prospects for further exploration and anticipates integrated LLM–RS systems for personalized services in varied scenarios.