PeopleofSuperintelligence

Researcher

Sahaj Agarwal

Also known as Sahaj Agrawal

Sahaj Agarwal is a research software engineer at Microsoft who co-authored papers on small and efficient language models, including Orca, Orca 2 and Phi-4-reasoning. AdaMix, a parameter-efficient tuning method he co-authored, appeared at EMNLP 2022. In 2023, he co-authored the Orca model, which learns from GPT-4 explanation traces, then SkipDecode, a method for faster inference, and Orca 2, which teaches small models different reasoning techniques. A 14-billion parameter reasoning model, Phi-4-reasoning, followed in April 2025, with him among the authors of its technical report. He is a Senior Research Software Engineer on the Microsoft Research AI Frontiers team, where Orca and Orca 2 are his featured work.

Works
9
Citations
237
h-index
4

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

  1. 2022

    The AdaMix tuning method

    In 2022, he co-authors AdaMix, a parameter-efficient method that tunes a mixture of adaptation modules in each Transformer layer while most pretrained-model weights stay frozen. The EMNLP 2022 paper, with Yaqing Wang as first author, reports that tuning 0.1–0.2% of the parameters outperforms full model fine-tuning on NLU and NLG tasks.

  2. June 2023

    Orca learns from GPT-4

    He is one of six authors of Orca: Progressive Learning from Complex Explanation Traces of GPT-4, posted in June 2023. The 13-billion parameter model learns to imitate the reasoning process of large foundation models, using GPT-4 explanation traces and step-by-step thought processes.

  3. July 2023

    SkipDecode for faster inference

    He co-authors SkipDecode in July 2023 with Luciano Del Corro, Allie Del Giorno, Bin Yu, Ahmed Awadallah and Subhabrata Mukherjee. The token-level early-exit method is built for batch inferencing and KV caching, and it obtains 2x to 5x inference speedups with negligible regression across a variety of tasks.

  4. November 2023

    Orca 2 for small models

    In November 2023, he co-authors Orca 2: Teaching Small Language Models How to Reason, which teaches small models several reasoning techniques. Orca 2 attains performance similar or better to models 5-10x larger on complex zero-shot reasoning tasks.

  5. April 2025

    The Phi-4-reasoning report

    In April 2025, he is a co-author of the Phi-4-reasoning Technical Report. Phi-4-reasoning is a 14-billion parameter reasoning model trained by supervised fine-tuning of Phi-4, and Phi-4-reasoning-plus adds a short phase of outcome-based reinforcement learning.

  6. As of October 2026

    Engineer at AI Frontiers

    He is a Senior Research Software Engineer on the Microsoft Research AI Frontiers team. Orca (June 2023) and Orca 2 (November 2023) are the work he features at Microsoft Research.

Most-cited works

  1. Application of ultrasonic vibration assisted MQL in grinding of Ti–6Al–4V2017 · Materials and Manufacturing Processes · 79 citations
  2. AdaMix: Mixture-of-Adaptations for Parameter-efficient Model Tuning2022 · Conference on Empirical Methods in Natural Language Processing (EMNLP) · 77 citationsWith Ahmed Awadallah and Jianfeng Gao
  3. Orca: Progressive Learning from Complex Explanation Traces of GPT-42023 · 68 citationsWith Ahmed Awadallah, Ganesh Jawahar and Hamid Palangi
  4. SkipDecode: Autoregressive Skip Decoding with Batching and Caching for Efficient LLM Inference2023 · 7 citationsWith Ahmed Awadallah
  5. Development of an ultrasonic vibration assisted minimum quantity lubrication system for Ti-6Al-4V grinding2019 · International Journal of Precision Technology · 2 citations

From OpenAlex, as of 21 October 2026.

  1. 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 Microsoft. 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.
  2. Google had the most people shaping AI every year from 2014 to 2025Includes Microsoft. Our estimates of which organizations employed the most people shaping AI each year from 1955 to 2025, tracing the shift from university labs to companies.
  3. The company share of AI researchers’ top papers was 31% in 2020–2024 and 8% in 1970–1974Includes Microsoft. 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. Google has the highest combined h-index among AI organizationsIncludes Microsoft. We estimate that Google holds the highest combined h-index at 8,016, while universities account for 73% of the deduplicated total across AI organizations.

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 Sahaj Agarwal yet. Send us a free-licensed photo.

  1. AdaMix: Mixture-of-Adaptations for Parameter-efficient Model Tuning - ACL Anthology aclanthology.org. Checked 20 October 2026.
  2. [2306.02707] Orca: Progressive Learning from Complex Explanation Traces of GPT-4 arxiv.org. Checked 20 October 2026.
  3. [2307.02628] SkipDecode: Autoregressive Skip Decoding with Batching and Caching for Efficient LLM Inference arxiv.org. Checked 20 October 2026.
  4. [2311.11045] Orca 2: Teaching Small Language Models How to Reason arxiv.org. Checked 20 October 2026.
  5. [2504.21318] Phi-4-reasoning Technical Report arxiv.org. Checked 20 October 2026.
  6. Sahaj Agarwal at Microsoft Research microsoft.com. Checked 20 October 2026.
  7. Microsoft Research AI Frontiers - Microsoft Research: People microsoft.com. Checked 20 October 2026.