WeFair-2B Query Rewriter Lifts CTR by 2.951% in WeChat Short-Video Search

Short-video search lives on ambiguous queries: a user types two words, and what they want to see depends on their recent searches and views. A team at East China Normal University has described WeFair, a way to train a compact 2B-parameter model that rewrites a query using the user's context without replacing the user's goal. The WeFair-2B model already runs in WeChat short-video search and serves more than 100 million queries per day. In a 10-day A/B test, relative CTR rose 2.951%, the share of effective video views rose 2.593%, and the share of query reformulations fell 2.350%.

The problem: engagement-based ranking pulls away from the query

Previous approaches distill the reasoning of large models into a small rewriter, then fine-tune it on downstream content-consumption signals: what the user watched after seeing the results. The authors point to a side effect. This kind of reward pushes the model toward queries that surface "sticky" content rather than what the person was looking for. This is called intent drift. In the paper's example, the query "crepe cake" drifts toward recipes, although the user wanted to watch people eating the cake. The rewritten query still draws on the user's real history, so formal checks have trouble catching this kind of error.

Large models are a poor fit for the task for two reasons. They understand the varied intents of a specific platform's users less well, and the latency of each call is too high for real-time search. In production, then, you need a small model that also decides whether to rewrite a query or leave it as is.

Three stages with different signal strictness

The idea behind WeFair is to match each training stage to its own level of feedback reliability.

• Demonstrations (SFT). Training starts with "context → query" records, then moves to reasoning over the context, then to directly predicting whether to reformulate or keep the query. A less precise but broader signal is acceptable here.
• Preferences (DPO). Pairs of candidates are distinguished by downstream consumption, and the model learns to choose those associated with a better hit on personal intent.
• Verification (GRPO). The reward is given only for an exact match with the user's actual reformulation, or with the decision to leave the query unchanged. The authors add no graded similarity scores, retrieval rewards, or judge-based rewards.

Further analysis shows that broad supervision helps at the SFT stage, while strict verification at the final stage yields better quality than the auxiliary rewards the authors tested. According to the authors, finer-grained signals reduce rewriting accuracy. At inference, the model immediately outputs a short rewrite or a "keep" marker, without generating a chain of reasoning, which saves latency.

How the training data was collected

The data comes from anonymized WeChat search sessions. For the "context → query" pool, the team selects queries with at least one effective video view (longer than ten seconds) and checks them with DeepSeek-V4-Flash for support from the context. For the reformulation pool, it links adjacent queries within one session: the first produced no video click, and the second produced an effective view. Queries triggered by system suggestions are excluded.

A two-stage filter follows. Rules remove additions made of Chinese function words, pairs with little character overlap (by Jaccard similarity), and cases with no contextual grounding. This eliminates about half of the candidates. Then three reasoning LLMs vote on intent continuity and information gain, after which three more check the "context bridge" between the added detail and the query. Roughly a third pass. Examples rejected at the last step become hard negatives for the keep decision, which teaches the model not to invent personalization without grounding.

What to keep in mind

The results come from WeChat's internal platform and cannot be independently reproduced. Part of the pipeline depends on Chinese-language rules and on closed behavioral logs. The source description includes no numerical offline-metric comparison with baseline models, only a statement of superiority in overall accuracy and intent preservation. The practical takeaway for teams building rewriters in search or recommendations: do not mix signals of different reliability at the same stage. Broad behavioral data works for the initial stage and for preferences, but the reward in the final RL step should be as clean as possible. The original query remains the anchor for retrieval and ranking, while the rewrite only expands the candidate pool.