A team of researchers from Tencent’s WeChat division has developed a system called WeFair that tackles a persistent challenge in short-video search: how to personalize query rewrites while staying faithful to the user’s original intent. The problem is pressing because users often type ambiguous or incomplete queries, forcing search engines to infer meaning quickly without misdirecting the results. Traditional approaches lean on synthetic reasoning or indirect behavioral rewards, which may produce more engaging content but risk drifting from what the user actually wanted. WeFair takes a different route by grounding its rewrites in real user behavior—specifically, the reformulations users type after seeing irrelevant results and the content they actually click and watch.
The system is built in stages. First, it mines actual human reformulations from search sessions, capturing the precise ways users adjust their queries when initial results fail. Then, it applies a multi-layered screening process—combining LLM-based checks and contextual verification—to filter out unsupported transitions, ensuring that rewritten queries don’t fabricate or distort intent. Training begins with a supervised curriculum using large-scale human demonstrations of personalized rewrites, gradually building the model’s capability. Next, it uses direct preference optimization to align with subsequent user behavior, such as what they watch after reformulating. Finally, an on-policy optimization stage introduces a strict verification reward based on real human reformulations, reinforcing fidelity to the user’s original need rather than just engagement metrics.
Critically, the researchers found that auxiliary rewards meant to boost exploration in group relative policy optimization (GRPO) often came at the cost of intent fidelity. The team’s online A/B testing in WeChat’s short-video search environment confirmed that WeFair delivers measurable gains: a 1.576% relative increase in effective clicked-query rate (defined as videos watched for more than 10 seconds) and a 1.731% relative drop in query reformulation rate. These numbers suggest users are finding relevant content faster, reducing the need to retry their searches.
While developed for a social media platform, the implications extend beyond consumer search. The fidelity-aware alignment framework offers a template for building trustworthy AI systems in contexts where precision matters—such as defense and security applications. Imagine a military or intelligence analyst querying a classified media archive with a vague request. A system that personalizes the query to the user’s role and past behavior must not invent details or shift intent, lest it produce irrelevant or misleading results. WeFair’s emphasis on grounding rewrites in real user corrections and verified transitions could help build safer, more reliable AI assistants in sensitive domains. The same fidelity-first approach—grounding outputs in observable human behavior rather than synthetic rewards—could reduce errors in high-stakes decision support systems where misinterpretation carries real consequences. Read the original research paper here.