Generative Recommendations in Recommender Systems

• Generative Recommendations (GRs) are systems that synthesize recommendations by directly modeling conditional output distributions with LLMs and diffusion architectures.
• They integrate content, behavior, and structure augmentation to unify multi-domain, multi-task, and multi-modal signals for personalized and explainable outcomes.
• GRs exploit scaling laws and emergent abilities to improve creative generation and transparency while addressing challenges like exposure bias and fairness.

Generative Recommendations (GRs) represent a paradigmatic shift in recommender systems, reformulating the core problem of matching users and items as a conditional generation task. Rather than ranking candidates via discriminative scoring, GRs directly model the conditional probability distribution over outputs—be those item identifiers, textual explanations, or interaction trajectories—conditioned on user context and interaction history. This synthesis-centric approach leverages the generative capabilities of LLMs and diffusion-based architectures, enabling unified, end-to-end modeling that capitalizes on world knowledge, natural language understanding, reasoning, and emergent creative abilities (Hou et al., 31 Oct 2025). The following sections provide an authoritative, technically rigorous overview of the GR domain as of 2026.

1. Generative vs. Discriminative Paradigms

Traditional recommender systems rely on discriminative models that estimate a scoring or relevance function, for example, $f(u, i) \approx P(y = 1 \,|\, u, i)$, and rank items accordingly. Model training optimizes objective functions such as the negative log-likelihood of observed labels, $$\mathcal{L}_{\text{disc}}(\theta) = -\sum_{(u, i, y)} \log \sigma(f_\theta(u, i))$$, where $\sigma$ denotes the sigmoid (Hou et al., 31 Oct 2025, Yang et al., 9 Jul 2025).

In contrast, generative recommendation models directly parameterize conditional generative distributions, e.g., $P(i\,|\,u)$ for item recommendation, or $P(\text{output}\,|\,u, \text{context})$ for richer outputs such as explanations or conversational responses. Training maximizes the likelihood of the ground-truth output, for example: $$\mathcal{L}_{\text{gen}}(\theta) = -\sum_{(u, i^+)} \log p_\theta(i^+\,|\,u)$$ or, for explanation generation,

$$\mathcal{L}_{\text{gen}}(\theta) = -\sum_{(u,\,\text{text}^+)} \log p_\theta(\text{text}^+\,|\,u)$$

This shifts the focus from item ranking to direct conditional synthesis, enabling new forms of user interaction, broader generalization, and the exploitation of foundation model scaling laws (Hou et al., 31 Oct 2025, Liu et al., 29 Sep 2025, Yang et al., 9 Jul 2025).

2. Data-Level Augmentation and Unification

2.1 Knowledge-Infused Augmentation

Generative models support a rich array of data augmentation strategies:

• Content Augmentation: Enriching user and item profiles with LLM-generated text (e.g., attribute summaries) to combat representation sparsity.
• Behavior Augmentation: Synthesizing pseudo-interactions, particularly for cold-start users, via generative sampling $\,\hat Y \sim G_\theta(\hat Y\,|\,u)\,$.
• Structure Augmentation: Extending or completing graphs (e.g., social or knowledge graphs) using LLMs to infer missing relations.

2.2 Agent-Based Simulation

LLM agents enable simulation of multi-turn user interactions and social network effects, generating large-scale, realistic logs to improve model robustness and evaluation (Hou et al., 31 Oct 2025).

2.3 Data Unification

GRs support various unification axes:

• Multi-Domain: Aligning heterogeneous signals (text, behavioral, cross-domain) into shared semantic or embedding spaces.
• Multi-Task: Reformulating diverse tasks (ranking, rating, explanation) under a common generation framework (e.g., text-to-text, as in P5).
• Multi-Modal: Unifying text, image, and ID modalities at the prompt or embedding level (e.g., Rec-GPT4V, UniMP), supporting universal or “one-model-for-all” approaches.

3. Model-Level Taxonomy and Architectures

3.1 LLM-Based Generative RS

• Prompted LLMs: Zero/few-shot recommendation via LLM prompting without task-specific fine-tuning.
• Aligned LLMs: Fine-tuned or LoRA-adapted LLMs on recommendation data, incorporating item-tokenization (assigning items to LLM vocabulary tokens) (Hou et al., 31 Oct 2025, Yang et al., 9 Jul 2025).
• Training Objectives: Include supervised fine-tuning (SFT), self-supervised contrastive learning, reinforcement learning (RLHF-style), and direct preference optimization (DPO):
  • $$\mathcal{L}_{\text{SFT}} = -\sum \log \pi_\theta(y^+|x)$$
  • $$\mathcal{L}_{\text{DPO}} = -\log \sigma\left[ \beta \left(\log \frac{\pi_\theta(y^+|x)}{\pi_\text{ref}(y^+|x)} - \log \frac{\pi_\theta(y^-|x)}{\pi_\text{ref}(y^-|x)} \right) \right]$$
• Inference: Direct generation with constrained decoding, retrieve-then-rerank pipelines, or acceleration via speculative decoding, distillation, and context truncation.

3.2 Large Recommendation Models (LRMs)

• HSTU: Treats entire user interaction sequences as autoregressive token streams, achieving predictable power-law scaling in quality as model and sequence size increase (Zhai et al., 2024, Chen et al., 1 Mar 2026).
• End-to-End Generative Retrieval & Ranking: Models such as OneRec, OneSug, and EGA-V2 encode history and context, then decode full ranked item lists, exposing a foundation model interface.

3.3 Diffusion-Based Generative RS

• Diffusion Data Augmentation: Generates high-quality pseudo-interactions and fills missing modalities via reversible noising/denoising steps (MoDiCF, DGFedRS).
• Generative Sampling: Treats target item vectors as noisy latents, denoising them conditionally on user context to produce final item embeddings (DiffRec, DreamRec).
• Modeling Diversity and Uncertainty: Guided diffusion to capture multifaceted preferences, supporting improved coverage and calibrated uncertainty (Hou et al., 31 Oct 2025).

4. Task-Level Capabilities and Applications

4.1 Core Tasks

• Next-Item and Top-K Generation: Conditional generation of ranked item lists $$\mathcal{L}_{\text{disc}}(\theta) = -\sum_{(u, i, y)} \log \sigma(f_\theta(u, i))$$0. Techniques include vocabulary-constrained decoding (e.g., FM-index, trie) and post-filtering.

4.2 Personalized Content Generation

• Visual: Virtual try-on, imagery synthesis (DreamVTON, OOTDiffusion).
• Textual: Personalized explanations, titles, reviews using LLM-prompting (PENS, PETER).

4.3 Conversational and Explainable Recommendation

• Conversational RS: GRs enable dialog-based recommendation via unified next-token modeling, supporting end-to-end optimization and structured interaction between intent detection and item generation. Retrieval-augmented architectures mitigate hallucination and allow grounding to external knowledge sources (Zhang et al., 21 May 2026).
• Explainability and Reasoning: Chain-of-thought (CoT) explanations and explicit reasoning traces are facilitated via prompt-driven generation or explicit reasoning tokens (ThinkRec, Reason4Rec), supporting transparent and trustworthy interfaces (Hou et al., 31 Oct 2025).

5. Advantages and Scaling Laws

GR approaches exhibit five primary advantages over discriminative RS paradigms (Hou et al., 31 Oct 2025):

1. World Knowledge Integration: Ability to inject encyclopedic and commonsense knowledge via pretrained LLMs.
2. Natural Language Understanding: Support for free-form user queries, reviews, and nuanced preference expression in both model input and output space.
3. Reasoning and Explainability: Structured generation enables explicit reasoning chains, causal inference, and interpretable recommendations.
4. Scaling Laws and Emergent Abilities: Model quality improves predictably with parameter and data scale, exposing emergent capabilities such as in-context learning and planning (Zhai et al., 2024, Liu et al., 29 Sep 2025).
5. Creative Generation: Unconstrained synthesis facilitates the creation of novel items, bundles, and explanations, moving beyond fixed candidate sets.

Scaling analyses reveal that while conventional SID-based GRs saturate due to discrete code capacity constraints, LLM-based GRs demonstrate continued improvements with increased model size, particularly when fine-tuned via adapters (e.g., LoRA); scaling both semantic information and collaborative filtering proficiency (Liu et al., 29 Sep 2025, Zhai et al., 2024).

6. Limitations, Open Challenges, and Future Directions

6.1 Evaluation and Benchmarks

• A need persists for dynamic, interactive benchmarks that capture multi-turn dialogs, multi-modal user feedback, and zero-shot generalization. Metric development for generative outputs—format robustness, hallucination, novelty, diversity, and fairness—is actively ongoing (Hou et al., 31 Oct 2025).

6.2 Robustness, Fairness, and Popularity Bias

• Exposure Bias: Standard SFT and DPO losses fail to explore plausible but unobserved items, leading to under-exposure of long-tail content. Solutions such as GFlowNets for trajectory-level learning tackle this directly (Wang et al., 19 Jun 2025).
• Popularity Bias: Token-level optimization and undifferentiated SID tokenization amplify exposure of head items. Techniques such as asymmetric unlikelihood optimization and structured tokenization (Ghost model) mitigate this at training and inference (Yin et al., 16 May 2026).
• Fairness and Debiasing: Prompt design, data augmentation, and adversarial defense are required to resist sensitive-attribute or position bias.

6.3 Efficiency and Deployment

• Fine-Tuning and Inference: Advances in fast alignment (coreset selection, continual pretraining), inference acceleration (speculative decoding), and training system design (jagged-acceleration, load balancing, asynchronous communication on specialized hardware) are now critical for real-world deployment (Chai et al., 13 May 2026).
• Industrial Pipelines: Integration of generative retrieval+ranking into real-time, log-latency pipelines is realized in large-scale advertising and content platforms, leveraging unified models for diverse business objectives (Xue et al., 26 Feb 2026).

6.4 Model Editing and Cold Start

• Efficient model editing—injecting new item knowledge in sub-training-epoch time via localized updates—addresses cold start collapse, allowing rapid real-world adaptation without retraining overhead (Shen et al., 15 Mar 2026).

6.5 Foundation Model Roadmap

• The field is converging on a vision of universal, multi-modal, multi-task recommendation assistants, leveraging unified LLM architectures, scalable training, and end-to-end generative pipelines for retrieval, ranking, explanation, and dialogue (Hou et al., 31 Oct 2025).

7. Outlook and Summary Table

Table: Model-Level Approaches in Generative Recommendation Systems.

Paradigm	Key Mechanism	Scaling/Capabilities
LLM-based GR	Prompting, Adapter Fine-Tuning	Predictable scaling, emergent abilities, reasoning
SID-based GR	Quantized Token Generation	Fast inference, limited by codebook capacity
Diffusion-based GR	Denoising, Temporal Masking	Captures diversity, uncertainty; temporal adaptation

GRs fundamentally reconceptualize recommendation as synthesis, unifying enhancements at the data, model, and task levels. This paradigm leverages LLMs’ world knowledge, unified representation learning, and scaling laws to provide generative, transparent, and interactive experiences, forming the architectural bedrock for next-generation recommendation assistants (Hou et al., 31 Oct 2025).