TwiSTAR: A Fast-Slow Adaptive Reasoning Framework for Generative Recommendation

This article introduces the TwiSTAR framework, which addresses the limitations of fixed reasoning strategies in generative recommendation. By using a fast-slow adaptive reasoning strategy, it significantly reduces inference latency while maintaining recommendation accuracy. The core of the framework is to adaptively allocate reasoning efforts for each user sequence, combining fast retrieval, lightweight ranking, and slow reasoning tools, with intelligent decision-making by a planner trained via reinforcement learning, providing new insights for efficiency optimization in recommendation systems.

Generative recommendation based on semantic IDs is a new paradigm, but existing methods use fixed reasoning strategies (either fast direct generation or slow chain-of-thought reasoning), leading to a dilemma: fast models have suboptimal accuracy on difficult samples; slow models have high latency and waste resources on simple cases. Balancing accuracy and efficiency has become a key challenge.

The core of the TwiSTAR framework is adaptive allocation of reasoning efforts, equipped with three main tools: 1. Fast SID Retriever (millisecond-level recall, suitable for simple scenarios); 2. Lightweight Candidate Ranker (quickly filters irrelevant items); 3. Slow Reasoning Model (generates natural language justifications, handles complex intents). A key innovation is converting inter-item collaborative knowledge into natural language explanations and injecting them into the slow model. The framework's core planner undergoes two-stage training (supervised warm-up + reinforcement learning) and dynamically decides which tool to call based on factors like user history complexity and candidate confidence.

Evaluations on three public datasets show: TwiSTAR consistently improves accuracy compared to fixed-strategy baselines (especially on difficult samples) while maintaining efficiency for simple samples; it significantly reduces latency and resource consumption compared to unified slow reasoning. Comparing with baselines: unified fast (low accuracy, extremely low latency), unified slow (high accuracy, extremely high latency), TwiSTAR achieves high accuracy + moderate latency/resource consumption.

Technical contributions include: first introduction of adaptive reasoning paradigm for generative recommendation; design of fast-slow three-layer architecture; collaborative knowledge injection to enhance reasoning; two-stage agentic training strategy. Application value: for platforms (cost reduction, support for large-scale requests, interpretability); for users (faster response, high-quality recommendations, transparent justifications); for businesses (reduced operational costs, improved retention, support for real-time scenarios).

Current limitations: cross-domain generalization ability of the planner needs verification; fixed tool design; offline training; only optimizes accuracy and latency. Future directions: improve planner generalization; explore fine-grained tool combinations; online learning to adapt to data changes; incorporate multi-objective optimization such as diversity.