Tingqiao Xu

Tingqiao Xu, Shaowei Yao, Chenhe Dong et al.

Accurately estimating query-item relevance is vital for e-commerce ranking and conversion. While Large Language Models (LLMs) excel at reasoning, they often lack specialized knowledge required for long-tail or fast-evolving queries, necessitating Retrieval-Augmented Generation (RAG). However, production environments face three critical challenges: (1) external context is inherently noisy and inconsistent; (2) extreme latency budgets prohibit multi-stage processing or refinement; and (3) the model must simultaneously assess relevance and context-trust within a unified inference pass. We propose DyKnow-RAG, a reinforcement learning framework that teaches LLMs to learn to trust through dynamic utilization of external knowledge. Built on Group Relative Policy Optimization (GRPO), DyKnow-RAG utilizes a dual-group rollout strategy (parametric-only vs. with-context) and a posterior-driven inter-group advantage scaling mechanism. This enables the model to optimize context utilization without human process labels or extra inference overhead. Our pipeline further integrates structured Chain-of-Thought (CoT) and an uncertainty-prioritized RL pool to stabilize training.Offline evaluations show significant Macro-F1 and Accuracy gains, particularly on noise-sensitive query slices. Importantly, DyKnow-RAG has been deployed in Taobao's production system, serving hundreds of millions of active users and billions of daily search requests. Controlled A/B tests demonstrate consistent lifts in key business metrics, including GSB and Item Goodrate, while maintaining a p99 latency under 400ms. This work provides a scalable and deployable paradigm for operationalizing noisy RAG under extreme efficiency constraints of large-scale industrial search.

Tingqiao Xu, Ziru Zeng, Jiayu Chen

The quality of supervised fine-tuning (SFT) data is crucial for the performance of large multimodal models (LMMs), yet current data enhancement methods often suffer from factual errors and hallucinations due to inadequate visual perception. To address this challenge, we propose VERITAS, a pipeline that systematically integrates vision priors and multiple state-of-the-art LMMs with statistical methods to enhance SFT data quality. VERITAS leverages visual recognition models (RAM++) and OCR systems (PP-OCRv4) to extract structured vision priors, which are combined with images, questions, and answers. Three LMMs (GPT-4o, Gemini-2.5-Pro, Doubao-1.5-pro) evaluate the original answers, providing critique rationales and scores that are statistically fused into a high-confidence consensus score serving as ground truth. Using this consensus, we train a lightweight critic model via Group Relative Policy Optimization (GRPO), enhancing reasoning capabilities efficiently. Each LMM then refines the original answers based on the critiques, generating new candidate answers; we select the highest-scoring one as the final refined answer. Experiments across six multimodal benchmarks demonstrate that models fine-tuned with data processed by VERITAS consistently outperform those using raw data, particularly in text-rich and fine-grained reasoning tasks. Our critic model exhibits enhanced capability comparable to state-of-the-art LMMs while being significantly more efficient. We release our pipeline, datasets, and model checkpoints to advance research in multimodal data optimization.