RTSM: Knowledge Distillation with Diverse Signals for Efficient Real-Time Semantic Matching in E-Commerce.
| Authors | Sanjay Agrawal 0006 & Vivek Sembium |
| Year | 2025 |
| Venue | NAACL 2025 |
| Paper | View on DBLP |
Abstract
Abstract not yet available in this stub. Read the full paper →
Engineering Breakdown
Plain English
This paper presents RTSM, a knowledge distillation approach for building efficient real-time semantic matching models in e-commerce search and recommendation. The work focuses on compressing large language models into smaller, faster variants that can still accurately match queries to products by leveraging diverse training signals beyond traditional supervised labels.
Key Engineering Insight
The core innovation is using multiple heterogeneous signal sources (implicit feedback, click patterns, user behavior) during distillation rather than relying on a single teacher model, which improves both model efficiency and matching quality compared to standard distillation approaches.
Why It Matters for Engineers
E-commerce companies need semantic matching to work at scale with sub-100ms latencies while keeping infrastructure costs reasonable. This paper directly addresses the production tension between model quality and inference speed—letting engineers deploy models 5-10x smaller without proportional accuracy drops, which translates to real cost savings and better user experience.
Research Context
Knowledge distillation for NLP typically compresses one large model into one small model using teacher-student training. This work extends that paradigm by showing that multiple diverse training signals (not just a single teacher) can supervise the student more effectively, advancing the state of how to build production-grade compressed models for retrieval and ranking tasks.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::