WeChat Releases WeFair: A Fidelity-Aware Personalized Query Rewriting System for Short-Video Search
Summary:WeChat has released WeFair, a fidelity-aware personalized query rewriting system for short-video search. WeFair employs a staged alignment strategy—demonstration, preference, and verification—to minimize intent drift and improve the accuracy of query rewriting based on user feedback. In offline experiments, WeFair-2B outperformed strong baselines in overall accuracy and intent preservation. Online A/B testing showed a 2.593% relative increase in the effective video-view rate and a 2.350% relativ
Technical Mechanism Analysis
WeFair implements personalized query rewriting through the following three stages:
-
Demonstration Stage: The model undergoes supervised fine-tuning, transitioning from large-scale context-to-query learning to direct prediction based on human rewrites. This stage aims to teach the model the semantic associations and intent expressions in user queries.
-
Preference Stage: Utilizing Direct Preference Optimization (DPO), the system learns relative preferences from subsequent consumption feedback to further enhance the accuracy of query rewriting.
-
Verification Stage: Through Group Relative Policy Optimization, the system rewards results that exactly match the rewrite or preservation targets, thereby constraining intent drift.
Engineering Trade-offs and Performance
- Strengths: WeFair's staged alignment strategy significantly reduces the intent drift problem common in traditional LLM-based methods. Its feedback fidelity assignment mechanism ensures precise capture of user intent.
- Challenges: The model may still need to rely on contextual information for inference when dealing with extremely short or highly ambiguous queries, which poses challenges for its generalization capabilities.
- Performance: In offline experiments, WeFair-2B outperformed strong baselines in overall accuracy and intent preservation. Online A/B testing showed a 2.593% relative increase in the effective video-view rate and a 2.350% relative reduction in the query reformulation rate.
Developer Deployment and Implementation Recommendations
- Deployment Environment: It is recommended to deploy WeFair in an environment with GPU acceleration to fully leverage the computational power of its deep learning models.
- Optimization Tips: Developers can adjust the feedback fidelity assignment mechanism according to specific application scenarios to achieve optimal performance. For instance, in scenarios where intent preservation is critical, increasing the weight of the verification stage is advisable.
- Future Directions: Future explorations could involve integrating WeFair with multimodal data (such as video content and user behavior data) to further enhance the search experience.
Conclusion
The release of WeFair marks a significant advancement in personalized query rewriting for short-video search. Its staged alignment strategy based on human feedback provides a new technical path for AI-driven search systems.
— END —Source: Hugging Face Daily Papers (2026-10-30)
Tags: #WeChat #Short-Video Search #Personalized Query Rewriting #Human Feedback
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments