TLDR: MiniOneRec is the first fully open-source framework for generative recommendation, demonstrating that larger models consistently improve performance on public datasets. It leverages Semantic IDs (SIDs) and a two-stage post-training process involving supervised fine-tuning and reinforcement learning. This process incorporates novel full-process SID alignment and reinforced preference optimization with hybrid rewards, leading to superior ranking accuracy and candidate diversity. MiniOneRec outperforms existing methods, shows strong transferability across domains, and benefits significantly from pre-trained Large Language Models, offering an efficient and effective solution for next-generation recommenders.
The world of recommendation systems is undergoing a significant transformation, driven by the remarkable success of large language models (LLMs). For years, traditional recommenders relied heavily on massive embedding tables to store user and item information. While effective to a point, this approach often hit a performance ceiling as these embedding dimensions grew.
A new paradigm, known as generative recommendation, is changing this. Instead of vast embedding tables, it uses compact Semantic ID (SID) sequences generated by autoregressive Transformers. This shift promises similar scaling benefits seen in LLMs, where increasing model size leads to consistent performance gains.
However, many industrial deployments of this generative approach have remained proprietary, leaving key questions unanswered for the broader research community: Do these scaling advantages hold true on public benchmarks? And what’s the simplest post-training method to achieve competitive performance?
Introducing MiniOneRec: An Open-Source Breakthrough
To address these questions, researchers have introduced MiniOneRec, the first fully open-source framework for generative recommendation. MiniOneRec provides a complete workflow, from constructing SIDs to supervised fine-tuning and recommendation-oriented reinforcement learning. The framework utilizes a Residual Quantized VAE to generate SIDs and then post-trains Qwen backbones, ranging from 0.5 billion to 7 billion parameters, on the Amazon Review dataset.
The experiments with MiniOneRec have yielded compelling results. They show a consistent decrease in both training and evaluation losses as the model size increases, validating the parameter efficiency of the generative approach. This means that, like LLMs, larger generative recommendation models tend to perform better.
Optimizing Performance with a Lightweight Pipeline
To further boost performance, MiniOneRec proposes a lightweight yet highly effective post-training pipeline. This pipeline focuses on two main areas:
-
Full-Process SID Alignment: This ensures that the model deeply understands the relationship between natural language and the discrete SID codes. It augments the vocabulary with dedicated SID tokens and enforces auxiliary alignment objectives throughout both the supervised fine-tuning (SFT) and reinforcement learning (RL) stages. This grounding in ‘world knowledge’ from LLMs is crucial for better recommendation accuracy.
-
Reinforced Preference Optimization: During the RL phase, MiniOneRec refines the generation process. It uses constrained decoding to ensure only valid items are produced, employs beam search for efficient exploration of diverse candidate recommendations, and utilizes a hybrid reward system. This reward system combines rule-based accuracy with a ranking-aware penalty, pushing the model away from less relevant items and improving both ranking accuracy and candidate diversity.
Also Read:
- Autonomous Reasoning for Smarter Recommendations: Introducing RecZero and RecOne
- MICRec: A Unified Framework for Smarter, More Adaptable Recommendations
Key Findings and Impact
MiniOneRec consistently outperforms traditional sequential recommenders, other SID-based generators, and even recent LLM-driven systems across various metrics on the Amazon Review dataset. Its ability to operate in a compact SID space also means faster inference, lower latency, and smaller memory footprints, making it highly efficient for real-world applications.
The framework also demonstrates strong transferability. When trained on one domain (e.g., Industrial) and deployed to an entirely unseen domain (e.g., Office) without further tuning, MiniOneRec can still uncover reusable interaction patterns, highlighting its potential for cross-domain recommendation.
Furthermore, the research confirms the significant impact of pre-trained LLMs. Models initialized with general-purpose pre-trained LLM weights consistently outperform those trained from scratch. This suggests that the reasoning ability and factual knowledge acquired during large-scale language pre-training provide a substantial advantage in understanding item semantics and discovering recommendation patterns.
MiniOneRec represents a significant step forward in making generative recommendation accessible and effective. It provides a robust, open-source platform for researchers and practitioners to explore and build the next generation of recommendation models. The codebase is publicly available on GitHub and Hugging Face, and future updates are planned to expand its capabilities. You can read the full research paper here: MiniOneRec: An Open-Source Framework for Scaling Generative Recommendation.
