LaRec: Unleashing LLM-based Latent Reasoning for Generative Recommendation

Traditional generative recommendation models often face a trade-off between the high-quality, step-by-step reasoning of explicit Chain-of-Thought (CoT) and the inference efficiency of latent reasoning. The authors investigate how to overcome the limitations of latent reasoning—specifically the lack of fine-grained supervision and the tendency toward deterministic, single-path reasoning—to create a more effective and efficient recommendation framework.

The authors propose LaRec, which employs a two-stage training paradigm:

1. Latent Pre-training (LPT): To provide supervision for hidden states, the model uses 'step-level alignment' to distill knowledge from explicit CoT reasoning paths generated by a teacher model. It also uses 'process direction alignment' to ensure that each latent reasoning step progressively moves the model closer to the target item in the embedding space.

2. Personalized RL-tuning (PRL): To address the deterministic nature of latent reasoning, the authors construct a personalized Gaussian Mixture Distribution for each user based on their history. During reinforcement learning, the model samples from this distribution to explore diverse reasoning paths within a user-relevant semantic space, rather than relying on random noise.

LaRec significantly outperforms existing latent reasoning baselines in recommendation accuracy while maintaining high inference efficiency. By anchoring exploration in personalized distributions and providing auxiliary supervision to the latent space, the model effectively captures multi-faceted user interests that deterministic latent models typically miss. The results demonstrate that the 'Pre-training then RL Alignment' paradigm is highly effective for scaling reasoning-enhanced recommendation systems.

This work provides a practical path for deploying reasoning-capable LLMs in latency-sensitive production environments. By moving reasoning into a continuous latent space and providing structured guidance, LaRec bridges the gap between the interpretability of explicit reasoning and the speed of traditional recommendation models.