Generative Recommendation with Large Language Models

75 concepts59 papers1 proposal

Traditional recommendation systems operate on a discriminative paradigm. Given a user's interaction history and a massive catalog of items, the system computes a matching score—often via a dot product of latent embeddings—for candidate items and sorts them. While highly efficient, this approach struggles to capture deep semantic context, perform multi-step reasoning, or adapt flexibly to open-world knowledge.

Generative recommendation fundamentally re-frames this problem. Instead of scoring a pre-defined list of candidates, a generative recommender treats recommendation as a conditional sequence generation task ($P(\text{next item} \mid \text{user history})$). By leveraging the autoregressive capabilities of Large Language Models (LLMs), these systems unify content understanding, user intent reasoning, and item retrieval into a single language modeling objective. As highlighted by Wenjie Wang 0007 et al. in The 2nd Workshop on Recommendation with Generative Models, this represents a paradigm shift away from traditional collaborative filtering toward context-aware, language-based reasoning.


Mapping Catalogs to Tokens: Semantic IDs and Vocabulary Expansion

To make an LLM act as a recommender, we must solve a fundamental representation problem: How do we represent a catalog of millions of items within a language model's token space?

If we represent items using their natural language titles, the model faces long sequence lengths, high inference latency, and the risk of hallucinating non-existent items. Conversely, if we assign arbitrary database IDs (e.g., Item_49201), the LLM's rich pre-trained semantic knowledge is completely wasted. Two primary strategies have emerged to resolve this tension:

  1. Semantic IDs (SIDs): Instead of random numbers, items are assigned structured, hierarchical identifiers that reflect their semantic properties (e.g., Books -> SciFi -> Novel_A). Neil Shah explores this in Generative Recommendation with Semantic IDs: A Practitioner's Handbook, demonstrating how structured identifiers allow generative models to navigate item spaces deterministically and leverage architectural choices like data augmentation effectively.
  2. Vocabulary Expansion: Rather than forcing items into existing text tokens, we can expand the LLM’s vocabulary so that every item in a catalog receives its own unique token. Haochen Zhang introduces this mechanism in CoVE: Compressed Vocabulary Expansion Makes Better LLM-based Recommender Systems. By mapping items directly to the model's logit output space and using hashing-based embedding compression, this approach transforms sequential recommendation into a high-speed next-token prediction task, bypassing the latency bottlenecks of traditional autoregressive decoding.

Conceptual Item Synthesis and Breaking the Filter Bubble

Discriminative recommenders are inherently constrained by the historical data they are trained on, often trapping users in "filter bubbles" by repeatedly recommending items highly similar to past clicks. Generative recommenders can break this cycle through a mechanism known as conceptual item synthesis.

Instead of scoring existing database entries, the model uses its open-world knowledge to generate a textual or conceptual description of an ideal item that a user might enjoy, even if that item does not exist or has never been interacted with in the training set.

[User History: Sci-Fi Movies & Baking Shows] 
       │
       ▼ (Traditional Discriminative Model)
[Scores existing Sci-Fi or Baking items independently]
       │
       ▼ (Generative Conceptual Synthesis)
"A futuristic cooking competition set in space" 
       │
       ▼ (Database Retrieval Match)
[Discovers highly serendipitous, cross-domain matches]

Muyan Weng et al. apply this concept in Bursting Filter Bubble: Enhancing Serendipity Recommendations with Aligned Large Language Models. By generating descriptive titles of optimal user-item pairings before matching them to a physical product database, their framework identifies unexpected yet highly relevant ("serendipitous") recommendations that traditional statistical models overlook.


Beyond Retrieval: Synthesizing Personalized Content

The generative paradigm extends beyond merely selecting an item; it allows the system to synthesize the entire user experience around that item. Traditional systems serve static assets (the same thumbnail and headline for every user). Generative recommendation pipelines use AI-Generated Content (AIGC) to tailor the presentation to individual user profiles.


LLMs as Central Reasoning Orchestrators

Deploying a massive language model to evaluate millions of items in real time is computationally prohibitive. To bridge the gap between heavy semantic reasoning and high-throughput industrial requirements, modern architectures position the LLM as a central reasoning orchestrator rather than a standalone retriever.

Instead of executing the entire recommendation pipeline internally, the LLM acts as an intelligent agent that coordinates specialized, traditional recommendation models as external tools.

┌────────────────────────────────────────────────────────┐
│               LLM Orchestration Engine                 │
│  - Infers complex user intent                          │
│  - Plans retrieval strategies                          │
└───────────────────┬────────────────────────┬───────────┘
                    │                        │
                    ▼ (Tool Call)            ▼ (Tool Call)
        ┌───────────────────────┐┌───────────────────────┐
        │ Traditional Retriever ││   Ranking & Filtering │
        │  (Vector DB / MIPS)   ││     (Graph Models)    │
        └───────────────────────┘└───────────────────────┘

For instance, Yuyue Zhao et al. introduce this paradigm in Let Me Do It For You: Towards LLM Empowered Recommendation via Tool Learning, where the LLM functions as a "surrogate user" that queries external tools over multiple rounds to explore item attributes. Similarly, the toolkit presented by Jianxun Lian et al. in RecAI: Leveraging Large Language Models for Next-Generation Recommender Systems delegates heavy-duty candidate generation and ranking to specialized models, using the LLM to provide conversational control, explainability, and interactive reasoning. This hybrid approach delivers the cognitive flexibility of generative models alongside the speed and reliability of traditional architectures.

Key Foundational Work

SoMin.ai: Personality-Driven Content Generation Platform (2020) — contributed 1 concept elements to this direction.
A Survey on Adversarial Recommender Systems: From Attack/Defense Strategies to Generative Adversarial Networks (2022) — contributed 1 concept elements to this direction.
DeepGen: Diverse Search Ad Generation and Real-Time Customization (2022) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_generative_recommendation_8d101ffc,
  title = {generative recommendation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generative-recommendation/rdr_8d101ffc46614dd6acb41c89e87f0a7b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_generative_recommendation_8d101ffc_v1,
  title = {generative recommendation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generative-recommendation/rdr_8d101ffc46614dd6acb41c89e87f0a7b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake