OneSearch-V2: The Latent Reasoning Enhanced Self-distillation Generative Search Framework
New generative search framework from 23 researchers delivers +3.98% click-through rates without extra latency.
A research team of 23 authors, led by Ben Chen, has unveiled OneSearch-V2, a major upgrade to their industrial-scale generative search framework. Generative Retrieval (GR) is a promising paradigm that offers end-to-end optimization and high efficiency compared to traditional multi-stage search systems. While the original OneSearch framework delivered commercial benefits, it struggled with complex queries, latent user intent, and overfitting to narrow historical data. The new V2 system directly tackles these limitations with three core technical innovations designed to make search smarter and more aligned with real user needs.
The first innovation is a thought-augmented complex query understanding module. This moves beyond simple keyword matching, enabling the system to perform deeper semantic reasoning to grasp the true intent behind ambiguous or complicated searches. The second is a reasoning-internalized self-distillation training pipeline. This technique allows the model to learn users' potential, precise intentions from data, going beyond just fitting to observed logs. The third is a behavior preference alignment optimization system, which helps mitigate 'reward hacking'—where a model optimizes for a single metric like conversions at the expense of overall quality—by incorporating direct user feedback.
The results from extensive testing are compelling. Offline evaluations confirmed strong improvements in query recognition and user profiling. More importantly, online A/B tests in a live e-commerce environment demonstrated clear business value: a +3.98% increase in item click-through rate (CTR), a +3.05% rise in buyer conversion rate, and a +2.11% boost in order volume. Manual evaluations also showed gains in search quality, with improvements in page good rate and query-item relevance. Crucially, OneSearch-V2 achieves these gains while mitigating common systemic issues like information bubbles and long-tail sparsity, and does so without adding any extra inference cost or serving latency, making it highly practical for deployment.
- Delivers +3.98% item CTR and +3.05% buyer conversion in live A/B tests, proving direct business impact.
- Uses a 'thought-augmented' module for deep query understanding, moving beyond shallow keyword matching.
- Solves systemic issues like information bubbles and long-tail sparsity without increasing inference costs or latency.
Why It Matters
This shows how next-gen AI search can directly boost e-commerce revenue and user experience, providing a blueprint for industrial deployment.