PeopleofSuperintelligence

Researcher

Hiroaki Hayashi

Hiroaki Hayashi is a research scientist at Salesforce AI Research who builds large language models for code and studies how language models behave in multi-turn conversation. He completed a PhD at Carnegie Mellon University's Language Technologies Institute in 2021 under Graham Neubig, after first-authoring the Eve optimization method in 2016, Latent Relation Language Models in 2019, and the WikiAsp summarization dataset in 2020. At Salesforce, he coauthored the CodeGen family of open code models in 2022 and 2023 and the XGen-7B technical report in 2023. His May 2025 paper "LLMs Get Lost In Multi-Turn Conversation", written with Philippe Laban, Yingbo Zhou and Jennifer Neville, found an average 39% performance drop in multi-turn settings and won an ICLR 2026 Outstanding Paper Award. He is a Senior Research Scientist at Salesforce AI Research, where he works on LLMs for code, interactivity in LLMs, and software engineering agents such as SAGE.

Badges

Timeline

  1. 2016–2019

    Eve optimizer and latent relation models

    In November 2016, he is first author of Eve, a gradient-based optimization method with locally and globally adaptive learning rates, written with Jayanth Koushik and Graham Neubig. With Zecong Hu, Chenyan Xiong and Graham Neubig, he publishes Latent Relation Language Models in August 2019.

  2. 2020–2021

    WikiAsp and a prompting survey

    In November 2020, he is first author of WikiAsp, a large-scale dataset for multi-domain aspect-based summarization built from Wikipedia articles in 20 domains, with Graham Neubig among the coauthors. He coauthors the survey "Pre-train, Prompt, and Predict" on prompting methods in natural language processing in July 2021 with Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang and Graham Neubig.

  3. 2021

    A Carnegie Mellon doctorate

    He completes a PhD at Carnegie Mellon University's Language Technologies Institute in 2021, advised by Graham Neubig, with a focus on natural language generation and the dissertation "Neural Aspect-based Text Generation". Upon graduation in 2021, he joins Salesforce as a researcher.

  4. 2022

    CodeGen, an open code model

    In March 2022, he is one of eight authors of CodeGen, a family of language models up to 16.1B parameters trained on natural language and programming language data for multi-turn program synthesis. The paper open-sources the JAXFORMER training library and introduces the Multi-Turn Programming Benchmark of 115 problem sets.

  5. 2023

    CodeGen2, CodeGen2.5 and XGen

    In May 2023, he coauthors CodeGen2, a study of training recipes for code and natural language models in sizes of 1B, 3.7B, 7B and 16B parameters. He shares equal contribution with Erik Nijkamp on the 7B-parameter CodeGen2.5 in July 2023, and in September 2023 he coauthors the XGen-7B technical report on 7B-parameter models trained with up to 8K sequence length for up to 1.5T tokens.

  6. May 2025

    Lost in multi-turn conversation

    With Philippe Laban, Yingbo Zhou and Jennifer Neville, he coauthors "LLMs Get Lost In Multi-Turn Conversation" in May 2025. Large-scale simulations in 2025 show that the top open- and closed-weight LLMs tested score an average of 39% lower in multi-turn conversations than in single-turn ones across six generation tasks.

  7. April 2026

    An ICLR Outstanding Paper

    The ICLR 2026 Outstanding Paper Committee names "LLMs Get Lost In Multi-Turn Conversation" one of two Outstanding Papers among 5,468 submissions for ICLR 2026, announced in April 2026. The committee points to its scalable method for evaluating multi-turn capabilities.

  8. As of October 2026

    Senior Research Scientist at Salesforce

    He is a Senior Research Scientist at Salesforce AI Research, working on LLMs for code, software engineering LLM agents, and interactivity in LLMs. In October 2025, he and eight coauthors introduce SAGE, a software engineering agent that first explores a problem and then induces an abstract plan from its own past experience.

Career

  1. NowSalesforceSenior Research Scientist

Education

  1. 2021Carnegie Mellon UniversityPhD

From the biography.

  1. Google alumni have founded or led 133 other organizations, more than any other AI labIncludes Salesforce. We estimate that alumni of 17 major AI labs went on to found or lead 350 other organizations, including 226 they founded, and we trace 472 career moves directly between 23 labs since 2010.
  2. Among employers with 30+ full-time cases, OpenAI leads primary technical role H-1B pay at $310,000, and Amazon, Google and Microsoft file 57% of all casesIncludes Salesforce. Across 440,669 certified US H-1B filings from 44 employers (FY2020–FY2026), OpenAI leads primary technical role base pay at $310,000 and xAI leads overall base pay at $290,000.
  3. The company share of AI researchers’ top papers was 31% in 2020–2024 and 8% in 1970–1974Includes Salesforce. We estimate where the people shaping AI worked when their most-cited papers appeared: the company share was 31% in 2020–2024 and 8% in 1970–1974, and Google led all organizations in 2020–2024.
  4. Do lab leaders still publish? 98% have published since 2021, drawing 2.3 times the median recent citations of researchersIncludes Salesforce. Among 86 lab leaders with confident publication records, 84 (98%) published since 2021, earning a median of 8,260 recent citations compared with 3,538 for researchers.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Hiroaki Hayashi yet. Send us a free-licensed photo.

  1. [1611.01505] Eve: A Gradient Based Optimization Method with Locally and Globally Adaptive Learning Rates arxiv.org. Checked 22 October 2026.
  2. [1908.07690] Latent Relation Language Models arxiv.org. Checked 22 October 2026.
  3. [2011.07832] WikiAsp: A Dataset for Multi-domain Aspect-based Summarization arxiv.org. Checked 22 October 2026.
  4. [2107.13586] Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing arxiv.org. Checked 22 October 2026.
  5. LTI Ph.D. Dissertations: 2021-2017 - Language Technologies Institute - School of Computer Science - Carnegie Mellon University lti.cmu.edu. Checked 22 October 2026.
  6. Hiroaki Hayashi, Author - Salesforce salesforce.com. Checked 22 October 2026.
  7. [2203.13474] CodeGen: An Open Large Language Model for Code with Multi-Turn Program Synthesis arxiv.org. Checked 22 October 2026.
  8. [2305.02309] CodeGen2: Lessons for Training LLMs on Programming and Natural Languages arxiv.org. Checked 22 October 2026.
  9. CodeGen2.5: Small, but mighty - Salesforce salesforce.com. Checked 22 October 2026.
  10. [2309.03450] XGen-7B Technical Report arxiv.org. Checked 22 October 2026.
  11. [2505.06120] LLMs Get Lost In Multi-Turn Conversation arxiv.org. Checked 22 October 2026.
  12. Announcing the ICLR 2026 Outstanding Papers – ICLR Blog blog.iclr.cc. Checked 22 October 2026.
  13. Hiroaki Hayashi hiroakih.me. Checked 22 October 2026.
  14. Software Engineering Agent via Self-Abstraction from Grounded Experience - Salesforce salesforce.com. Checked 22 October 2026.