WeFair: Fidelity-Aware Alignment from Human Feedback for Personalized Query Rewriting in Short-Video Search

WeFair. Fidelity-Aware Alignment from Human Feedback for Personalized Query Rewriting in Short-Video Search. By Yingtao Ren and nine others, from Tencent's Weixin Group and East China Normal University. Published on the thirtieth of October.

Short-video search needs query rewriters that are compact, and that recover what a particular user actually means from a query that is ambiguous, all under strict latency constraints. Existing methods built on large language models have improved rewriting in two ways: with synthetic reasoning, and with feedback from what users go on to consume. That second signal carries a hazard. Someone who types a vague query and then watches an engaging clip has not necessarily found what they came for. Lean on subsequent consumption too heavily and the rewriter learns to favour engaging content over the user's immediate need. The result is intent drift, and it happens even when the rewritten query is grounded in the user's behavioural history.

WeFair, short for Fidelity-Aware Alignment from Human Feedback, is built for personalized query rewriting in WeChat short-video search. Its organising principle is that human feedback is not equally faithful to intent. Each kind is used according to how directly it reflects what the user personally wanted, and the kinds are arranged as a progression of fidelity: demonstration first, then preference, then verification. Demonstration shows the model what users wrote. Preference tells it what they went on to choose. Verification holds it to an exact target.

Training data comes from real search sessions. Actual queries and the reformulations users made are mined from those sessions, then screened for contextual support. In the demonstration stage, a supervised fine-tuning curriculum moves from large-scale context-to-query learning toward reasoning and direct prediction on human reformulation and on preservation.

The preference stage uses direct preference optimization to learn relative preferences from subsequent consumption feedback. The engagement signal still enters, then, but in the middle of the progression rather than at its end. Verification comes last. There, group relative policy optimization rewards exact agreement with a reformulation or preservation target, and that strictness is what constrains intent drift.

Further analysis separates two design choices. Broad supervision benefits the full curriculum of supervised training, and strict verification performs better than auxiliary rewards. Offline, WeFair-2B outperforms strong rewriting baselines in overall accuracy and in intent preservation. Online, an A/B test shows a relative increase of two point five nine three percent in effective video-view rate. Query reformulation rate, meanwhile, falls by a relative two point three five percent. Both online numbers point the same way: more views that count as effective, and fewer users needing to rephrase what they typed.

WeFair-2B is deployed in the WeChat short-video search system, which serves over one hundred million queries per day. Read as a whole, the design is an ordering of trust. The feedback that shows intent most directly is placed at the end of training, where it has the final word, while the signal most prone to drift is kept in the middle: informative, but not the last thing the model learns from.