Institution profile

Coupang

Industry researchasia · kr
Official website
Research library9linked papers
Opportunities0open roles
Selected work

Representative Papers

On the Practice of Scaling Search Conversion Rate Prediction

May 27, 2026

This work addresses the challenge of scaling conversion rate (CVR) prediction models in high-traffic scenarios, where trade-offs among model quality, training cost, and serving latency are critical. Through empirical analysis, the authors find that scaling effects from backbone architecture, embedding parameter size, and training data volume are approximately independent and additive. Leveraging this insight, they propose a lightweight warm-start strategy to accelerate training and integrate inference optimizations—including decoupled graph execution and dynamic batching—to enable low-latency GPU deployment of high-capacity models. In online A/B experiments with 2.5× more training data and 8× greater inference compute, the approach achieves a 2.6% improvement in key CVR metrics with negligible latency overhead.

0 citationsRead paper

CSRM-LLM: Embracing Multilingual LLMs for Cold-Start Relevance Matching in Emerging E-commerce Markets

Sep 01, 2025

To address the cold-start relevance matching challenge in emerging e-commerce markets—characterized by scarce tag and user behavioral data—this paper proposes the Cross-lingual Semantic Relevance Matching (CSRM) framework. First, it leverages machine translation as a pretraining task to activate cross-lingual transfer capabilities of multilingual large language models. Second, it incorporates a retrieval-augmented query understanding module to enable semantic-driven query expansion. Third, it introduces a multi-round self-distillation training strategy to mitigate annotation noise and enhance generalization under low-resource conditions. CSRM operates without human annotations and significantly reduces reliance on historical data from the target market. Online deployment results demonstrate a 45.8% reduction in system defect rate and a 0.866-percentage-point increase in session purchase rate, substantially improving search and recommendation quality in cold-start scenarios.

0 citationsRead paper

DualFit: A Two-Stage Virtual Try-On via Warping and Synthesis

Aug 16, 2025

Virtual try-on methods often fail to faithfully preserve high-frequency garment details (e.g., logos, prints), undermining brand representation and user trust. To address this, we propose DualFit—a two-stage high-fidelity virtual try-on framework. In Stage I, a learned optical flow-based deformation network achieves precise garment-to-body alignment. In Stage II, a region-aware inpainting mechanism—incorporating semantic preservation masks and diffusion-based image completion—is applied exclusively to non-critical regions, thereby strictly preserving high-frequency structural details (e.g., logos) in their original form. DualFit is the first approach to jointly model structured semantic preservation and generative synthesis, significantly enhancing detail integrity and visual naturalness. Extensive experiments demonstrate that DualFit substantially outperforms state-of-the-art methods across multiple benchmarks, particularly in logo/print fidelity, boundary blending quality, and overall perceptual realism.

0 citationsRead paper
Recent publications

Latest Papers

On the Practice of Scaling Search Conversion Rate Prediction

May 27, 2026

This work addresses the challenge of scaling conversion rate (CVR) prediction models in high-traffic scenarios, where trade-offs among model quality, training cost, and serving latency are critical. Through empirical analysis, the authors find that scaling effects from backbone architecture, embedding parameter size, and training data volume are approximately independent and additive. Leveraging this insight, they propose a lightweight warm-start strategy to accelerate training and integrate inference optimizations—including decoupled graph execution and dynamic batching—to enable low-latency GPU deployment of high-capacity models. In online A/B experiments with 2.5× more training data and 8× greater inference compute, the approach achieves a 2.6% improvement in key CVR metrics with negligible latency overhead.

0 citationsRead paper

CSRM-LLM: Embracing Multilingual LLMs for Cold-Start Relevance Matching in Emerging E-commerce Markets

Sep 01, 2025

To address the cold-start relevance matching challenge in emerging e-commerce markets—characterized by scarce tag and user behavioral data—this paper proposes the Cross-lingual Semantic Relevance Matching (CSRM) framework. First, it leverages machine translation as a pretraining task to activate cross-lingual transfer capabilities of multilingual large language models. Second, it incorporates a retrieval-augmented query understanding module to enable semantic-driven query expansion. Third, it introduces a multi-round self-distillation training strategy to mitigate annotation noise and enhance generalization under low-resource conditions. CSRM operates without human annotations and significantly reduces reliance on historical data from the target market. Online deployment results demonstrate a 45.8% reduction in system defect rate and a 0.866-percentage-point increase in session purchase rate, substantially improving search and recommendation quality in cold-start scenarios.

0 citationsRead paper

DualFit: A Two-Stage Virtual Try-On via Warping and Synthesis

Aug 16, 2025

Virtual try-on methods often fail to faithfully preserve high-frequency garment details (e.g., logos, prints), undermining brand representation and user trust. To address this, we propose DualFit—a two-stage high-fidelity virtual try-on framework. In Stage I, a learned optical flow-based deformation network achieves precise garment-to-body alignment. In Stage II, a region-aware inpainting mechanism—incorporating semantic preservation masks and diffusion-based image completion—is applied exclusively to non-critical regions, thereby strictly preserving high-frequency structural details (e.g., logos) in their original form. DualFit is the first approach to jointly model structured semantic preservation and generative synthesis, significantly enhancing detail integrity and visual naturalness. Extensive experiments demonstrate that DualFit substantially outperforms state-of-the-art methods across multiple benchmarks, particularly in logo/print fidelity, boundary blending quality, and overall perceptual realism.

0 citationsRead paper