BiFlowRec: Bidirectional Rectified Flow for LLM-Conditional Generative Recommendation

Generative recommendation models user preferences over the entire item set. With the rapid advances of large language models (LLMs), an emerging challenge is how to let LLM semanticsdirectly shapethis generative process rather than merely serving as side features. Existing approaches to LLM-conditioned generative recommendation typically perform distribution alignment in a VAE latent space, which introduces an encode–align–decode pipeline and can incur information loss and additional computational overhead. We propose BiFlowRec, aRectified Flow-based symmetric bidirectional conditional flow matching model that directly constructs LLM-conditioned vector fields in the interaction space for both user→item and item→user directions. BiFlowRec employs semantics-aware confidence weighting to suppress noisy interactions and leverages the straight-line property of Rectified Flow to support single-step generation. Experiments on three public datasets (Amazon-Book, Yelp, and Steam) show that BiFlowRec, as an end-to-end generative recommender, consistently outperforms a wide range of generative and LLM-enhanced baselines. In particular, BiFlowRec improves Recall@10 by 34.0%, 6.8%, and 2.3% on the three datasets, respectively, over the strongest generative baselines. Ablation studies further validate the effectiveness of the symmetric bidirectional architecture, the semantics-aware sample reweighting strategy, and the overall training design.