HoloRec embeds chain-of-thought reasoning into generative recommendation with two modes
New generative recommendation model builds chain-of-thought reasoning without expensive annotations.
HoloRec (Holistic Encoding and Interleaved Reasoning for Generative Recommendation) addresses two key limitations of existing generative recommenders: flat semantic representations that hinder multi-step reasoning, and the reliance on externally constructed chain-of-thought (CoT) data, which is expensive to annotate and disconnected from the generation objective. The model constructs a hierarchical semantic encoding matrix via multi-granularity nested residual quantization, optimized end-to-end with a holistic reconstruction loss. This unified representation allows reasoning to be embedded directly into the generation process.
HoloRec offers two modes: a non-thinking mode that uses lightweight multi-granularity supervised alignment for rapid predictions, and a thinking mode that employs an interleaved reasoning scheme to generate CoT steps without external data. Evaluated on multiple public recommendation datasets, HoloRec consistently outperforms baselines, with especially significant gains in sparse scenarios. The thinking mode yields better accuracy than the non-thinking mode while adding only modest inference overhead, making it practical for deployment.
- HoloRec uses endogenous chain-of-thought reasoning, eliminating the need for expensive external CoT annotations.
- Two inference modes: non-thinking (fast, lightweight) and thinking (interleaved reasoning, higher accuracy at modest overhead).
- Significant performance gains on sparse recommendation datasets, consistently outperforming baseline models.
Why It Matters
Brings structured reasoning to recommendation systems, improving accuracy without costly human annotations or external data.