Generative recommendation via semantic token prediction

What it does

MiniOneRec treats recommendation as next-token prediction. It compresses each product—title, description, and all—into a compact Semantic ID using a three-level RQ-VAE quantizer. An LLM then learns to continue a user’s chronologically ordered history, and a subsequent GRPO phase refines the policy with recommendation-specific rewards.

The interesting bit

Because the action space is a closed catalog of item tokens, the system uses constrained beam search to ensure every generated recommendation is a valid, unique product. That neatly avoids the classic generative-recommender failure mode of hallucinating items that do not exist.

Key highlights

• End-to-end open-source pipeline covering SID construction, supervised fine-tuning, and RL
• Semantic IDs derived from frozen text-encoder embeddings rather than arbitrary indices
• GRPO training with group-normalized rewards, a KL penalty, and rank-aware penalties for incorrect high-probability items
• Constrained decoding guarantees valid and unique beam outputs
• Targets heavy hardware: expects 4–8 A100/H100 GPUs for reproduction

Caveats

• Reproducibility rough edges exist: constrained decoding can break with Instruct-tuned models, producing invalid items; the authors suspect transformers library drift and currently recommend switching to base models like Qwen2.5-base
• The data pipeline and preprocessing scripts are tightly bound to Amazon review datasets (2018/2023), so bringing your own catalog requires writing new adapters

Verdict

A solid reference for researchers exploring LLM-based recommendation. Production teams without a rack of A100s should look elsewhere.