Skip to content
TheChatArchiveThe Chat Archive
ChatGPT (model unspecified)· OpenAIHugging Face ShareLM (browser-extension contributions)

Who is Leshem Choshen

  1. User

    Who is Leshem Choshen

  2. Assistant

    Leshem Choshen is a researcher in Natural Language Processing and machine learning, currently a postdoctoral researcher at MIT’s CSAIL and a researcher at IBM Research in Cambridge, MA YouTube+2IBM Research+2IBM Research+2. He completed his PhD at Hebrew University of Jerusalem (2017–2022) under Omri Abend, where he explored topics like collaborative pretraining, model evaluation, and neural model merging ResearchGate+2openreview.net+2dblp.org+2. He’s earned several academic honors including Rothschild Fellowship, being offered a Fulbright Fellowship, Blavatnik Award, and KLA scholarships ktilana.wixsite.com. Leshem has contributed to many prominent publications, including recent work such as A Hitchhiker’s Guide to Scaling Law Estimation (ICML 2025) and Compress then Serve: Serving Thousands of LoRA Adapters with Little Overhead (ICML 2025), along with projects like DOVE and models related to the BabyLM benchmark and model merging evaluations ktilana.wixsite.com+4IBM Research+4IBM Research+4.

    🔍 Research Focus Areas

    Model merging and collaborative pretraining (efficiently combining fine‑tuned models, learning across domains)

    Evaluation of large language models, including bias mitigation and benchmark design

    Scaling law estimation and efficient serving of models (e.g. LoRA adapter compression strategies)

    🏛️ Institutional Affiliations

    IBM Research – Cambridge, MA: has held a researcher position since around 2016 or earlier, contributing to AI and NLP research ktilana.wixsite.comgithub.com+2openreview.net+2IBM Research+2.

    MIT CSAIL: currently a postdoctoral researcher since 2024 openreview.netil.linkedin.com.

    Would you like to dive deeper into any specific research topic of Leshem Choshen—say, model merging techniques, the BabyLM benchmarks, or one of his recent papers?

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.