RecGPT-V3 Technical Report

Summary & Methodology Analysis

RecGPT-V3 builds upon the RecGPT-V1 and RecGPT-V2 systems on Taobao to address challenges where large language models in recommender systems suffer from stateless behavior modeling, a tag-to-item information bottleneck, and inefficient explicit reasoning at scale. The system utilizes a Global Planner and multi-expert sub-agents for intent analysis, alongside Qwen3-14B as the foundation backbone and CN-CLIP plus RQ-VAE for quantization. The paper introduces a Memory Hub that condenses the full behavioral sequence into structured memory units, maintains provenance through representative indices, and incrementally updates preferences through Evolving Memory Curation. Additionally, it employs a Hybrid-modal Foundation Model that extends the Qwen3-14B vocabulary with Semantic IDs alongside natural language, and trains it via continual pre-training and instruction tuning.

For reasoning and policy optimization, RecGPT-V3 implements Latent Intent Reasoning, which internalizes verbose explicit chains of thought into compact learnable latent tokens that can be decoded back into readable rationales on demand. The policy is optimized via Reinforcement Learning from Ranking Feedback using rewards read from the downstream production ranking model. These architectural choices allow structured behavior compression to achieve an 80% token reduction, cutting user-modeling computation by 55.8% and serving resource consumption by 52.4% in the feed scenario, while simultaneously improving IPV by +1.28%, CTR by +1.00%, TC by +1.97%, and GMV by +3.97%.

Despite these improvements, the paper notes a key limitation: training predominantly on Semantic ID grounding data risks catastrophic forgetting of general-purpose competence. The paper does not specify other limitations or hardware requirements beyond what is stated regarding this grounding data risk.