PulseAugur
EN
LIVE 20:21:41

HARNESS-LM trains compact models for faster sponsored search retrieval

Researchers have developed HARNESS-LM (HLM), a novel three-phase training framework designed to transfer the capabilities of large language models into compact, efficient models for sponsored search retrieval. This method involves training a high-performance "teacher" model, distilling its knowledge into a smaller "student" encoder, and then refining the student for optimal retrieval performance. HLM successfully recovers over 98% of the teacher model's precision while significantly reducing latency and increasing throughput, demonstrating practical efficacy through A/B testing on Bing Ads. AI

IMPACT Enables the deployment of powerful language models in latency-sensitive applications, improving efficiency and performance in areas like sponsored search.

RANK_REASON Publication of an academic paper detailing a new training framework for language models.

Read on arXiv cs.IR (Information Retrieval) →

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →

COVERAGE [2]

  1. arXiv cs.AI TIER_1 · Vipul Gupta, Shikhar Mohan, Lakshya Kumar, Pranjal Chitale, Nikit Begwani, Amit Singh, Manik Varma ·

    HARNESS-LM: A Three-Phase Training Recipe for Harnessing SLMs in Sponsored Search Retrieval

    arXiv:2605.23572v1 Announce Type: cross Abstract: In the competitive landscape of sponsored search, balancing retrieval quality with production latency is a critical challenge. While large retrieval models based on Small Language Models (SLMs) such as Qwen3-Embedding-4B/8B set st…

  2. arXiv cs.IR (Information Retrieval) TIER_1 · Manik Varma ·

    HARNESS-LM: A Three-Phase Training Recipe for Harnessing SLMs in Sponsored Search Retrieval

    In the competitive landscape of sponsored search, balancing retrieval quality with production latency is a critical challenge. While large retrieval models based on Small Language Models (SLMs) such as Qwen3-Embedding-4B/8B set strong upper bounds on public benchmarks, their depl…