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

WeFair trains a compact personalized short-video query rewriter by matching human-feedback fidelity to curriculum SFT, consumption-pair DPO, and strict-target GRPO, reducing intent drift and improving online engagement.

In WeChat short-video search, ambiguous queries must be rewritten or preserved under strict latency constraints, using query-time context to recover the user's immediate need. Existing LLM-based rewriters use synthetic reasoning and subsequent consumption feedback, but overemphasizing consumption can favor engaging content and cause intent drift. Recent methods use mined user reformulations, yet how directly different feedback reflects personalized intent and how to use it across training stages remains underexplored.

WeFair mines context-to-query, reformulation, and preservation pools from search sessions; reformulation candidates are rule-filtered and then screened by three-model majority votes for intent continuity, information gain, and contextual bridging. Curriculum SFT first predicts queries from context, then learns teacher rationales, then predicts human reformulation or preservation targets directly. DPO then uses candidate pairs distinguished by human reformulations or subsequent consumption, and GRPO rewards exact agreement with screened reformulation or preservation targets with no graded-similarity, retrieval, or judge rewards. At serving, rewrites expand cache and vector retrieval while the original query remains a ranking anchor, gated by the preservation-token probability.

• On offline personalized rewriting, WeFair-4B achieved the highest overall accuracy and WeFair-2B achieved the highest consistent recall among the compared methods. §4.2
• WeFair-2B achieved an intent drift rate of 5.36, lower than WeWrite-0.8B at 13.34 and AWHCP-2B at 14.88, and the paper describes this as about one-third of comparable rewriting methods. §4.2
• WeFair's overall accuracy increased by 5.91 percentage points from 0.8B to 2B but only 0.67 points from 2B to 4B. §4.2
• Strict verification in GRPO achieved the highest positive and overall advantage purity and the strongest overall accuracy among the tested reward variants. §4.4
• WeFair-2B direct decoding used 7.60 tokens and 0.316 s per sample, versus 26.65 tokens and 0.795 s for WeFair-CoT-2B; at 4B, direct WeFair used 72.66% fewer tokens and achieved a 2.32× speedup over its CoT variant. §4.5
• In a 10-day online A/B test, cache plus vector retrieval increased CTR by 2.951% and VV>10s by 2.593%, and reduced QRR by 2.350%, with P<0.01. §4.6

• The offline benchmark uses an LLM judge; human agreement was checked on a random sample of 500 Gemini-3.1-Pro responses only, leaving judge reliability across all systems and cases as an evaluation scope stated in the paper.
• Online evidence comes from one 10-day A/B test in WeChat short-video search; generalization to other platforms or domains is not tested.
• Preference construction relies on sensitive search and consumption histories, and the paper withholds raw user-level records and some internal feedback-processing details, limiting reproducibility.
• AWHCP is a local platform adaptation replacing e-commerce data and feedback services, so baseline comparisons reflect an adapted implementation.

• WeFair assigns feedback fidelity as demonstration, preference, and verification, with each training stage initializing the next. §3.2
• Training data comprised 1500K context-to-query, 138K human reformulation, and 461K preservation examples. §A.2
• Reformulation mining requires no click after q and an effective view after q', and rule filtering removes no character gain, function-word-only gains, Jaccard <= 0.06, and insufficient contextual support. §3.1
• DPO generated 64 candidates per input and constructed up to two pairs, requiring chosen top-10 retrieval to contain a relevant consumed video and rejected top-100 retrieval to contain none. §3.2
• GRPO sampled 32 responses per query and rewarded exact match with reformulation or preservation, giving zero reward to all other outcomes. §3.2
• Rewrite-based relevance is gated by the preservation-token probability; otherwise relevance is computed for both the original query and rewrite and combined in ranking. §3.3
• The full SFT curriculum achieved 55.48 overall accuracy versus 47.16 for HR Only. §4.3.1
• Production WeFair generates rewrites for 42.26% of processed queries and preservation markers for 57.74%, with P95 inference latency of 81 ms and +16 ms end-to-end search latency. §4.6