S$^2$GR: Stepwise Semantic-Guided Reasoning in Latent Space for Generative Recommendation

2026-01-26

AI Summary

This work addresses the limitations of existing generative recommendation methods, which lack deep, interpretable reasoning mechanisms when directly generating semantic IDs from interaction sequences and suffer from a disconnection between reasoning and generation. To overcome this, we propose the S²GR framework, which introduces a stepwise semantic guidance mechanism in the latent space by inserting interpretable “thought tokens” prior to each semantic ID generation, enabling fine-grained, semantically aligned reasoning. We innovatively incorporate contrastive learning to provide semantic supervision for these thought tokens and design an optimized codebook structure to enhance both coarse-to-fine semantic hierarchy and computational balance. Extensive experiments demonstrate that S²GR significantly outperforms state-of-the-art methods across multiple benchmarks, and online A/B tests on a large-scale short-video platform confirm its practical effectiveness in improving recommendation performance.

Technology Category

• Natural Language Processing: Generation
• Knowledge Representation and Reasoning: Common-Sense Reasoning
• Computer Vision: Visual Reasoning & Symbolic Representations

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

Abstract

Generative Recommendation (GR) has emerged as a transformative paradigm with its end-to-end generation advantages. However, existing GR methods primarily focus on direct Semantic ID (SID) generation from interaction sequences, failing to activate deeper reasoning capabilities analogous to those in large language models and thus limiting performance potential. We identify two critical limitations in current reasoning-enhanced GR approaches: (1) Strict sequential separation between reasoning and generation steps creates imbalanced computational focus across hierarchical SID codes, degrading quality for SID codes; (2) Generated reasoning vectors lack interpretable semantics, while reasoning paths suffer from unverifiable supervision. In this paper, we propose stepwise semantic-guided reasoning in latent space (S$^2$GR), a novel reasoning enhanced GR framework. First, we establish a robust semantic foundation via codebook optimization, integrating item co-occurrence relationship to capture behavioral patterns, and load balancing and uniformity objectives that maximize codebook utilization while reinforcing coarse-to-fine semantic hierarchies. Our core innovation introduces the stepwise reasoning mechanism inserting thinking tokens before each SID generation step, where each token explicitly represents coarse-grained semantics supervised via contrastive learning against ground-truth codebook cluster distributions ensuring physically grounded reasoning paths and balanced computational focus across all SID codes. Extensive experiments demonstrate the superiority of S$^2$GR, and online A/B test confirms efficacy on large-scale industrial short video platform.

Problem

• Generative Recommendation
• Semantic ID
• Reasoning Mechanism
• Latent Space
• Codebook

Innovation

• stepwise reasoning
• semantic-guided generation
• codebook optimization
• contrastive learning
• generative recommendation