Amir Keivan Mohtashami
I am a Ph.D. student at EPFL working on interested in understanding and improving capabilities of foundational models. Examples include finding better data formats and improved architectural choices through to allow the model perform inference at infinite context length in a practically feasible way, allowing the model to process input from multiple sources / speakers, and improving reasoning capabilities. I am also interested in improving the training stability and speed of these models by first understanding the reason for the success of current optimizers as well as the reason for the observed instabilities and using this understanding to develop faster and more stable optimizers.
- Identity CONFIRMED across independent sources: IOI stats page lists Amir Keivan Mohtashami (Iran, IOI 2015, silver medal) with Codeforces handle 'matrix' and TopCoder handle 'snbk'; the same person is 'matrix' on Codeforces, 'snbk' on TopCoder/CodeChef, and 'Amirkeivan Mohtashami' on OpenReview/EPFL with a 2020-2025 EPFL PhD and 2015-2020 Sharif undergrad matching the olympiad timeline.Jan 1, 2026
- Current role: Research Scientist at Google (Google DeepMind), Zurich, since 2025 - stated on his OpenReview profile (Research Scientist, Google, 2025-Present), LinkedIn headline, and X bio ('PhD - EPFL'); not inferred from an old index.Jan 1, 2025
- Quant employment verification: UNRESOLVED. The only quantitative-finance employment evidence is an index-derived, self-reported 'Quantitative Analyst Intern' entry at The D. E. Shaw Group (Jun-Aug 2024) surfaced via LinkedIn profile enrichment (https://linkedin.com/in/akmohtashami) - index-derived, not independent primary-source confirmation, so the role is not independently verified against a primary employer source. Employer classification: The D. E. Shaw Group is a multi-strategy quantitative hedge fund and does operate in quantitative finance, so IF the role is accurate it would be a genuine quant-finance internship; the classification alone does not verify that this specific role was held. No competitions, insight days, camps, scholarships, or student programs are counted as employment. All other listed roles (Google/Google DeepMind research, EPFL doctoral study, Sotoon data science, Sharif research assistantship) are non-quant-finance.Jun 1, 2024
- LinkedIn enrichment (index-derived, self-reported on his LinkedIn profile) lists a 'Quantitative Analyst Intern' role at The D. E. Shaw Group; D.E. Shaw Group is a multi-strategy quantitative hedge fund. See the 'Quant employment verification' note for the verdict.Jun 1, 2024
- IOI profile (stats.ioinformatics.org/people/5617) confirms Amir Keivan Mohtashami represented Iran at IOI, with Codeforces handle 'matrix' and TopCoder handle 'snbk'; listed as ITC member for Iran 2023-2025 and Switzerland 2026.Jan 1, 2026
- Education timeline: undergrad at Sharif University of Technology 2015-2020, PhD in Computer Science at EPFL 2020-2025; OpenReview career history lists Sharif undergrad 2015-2020 and EPFL PhD 2020-2025.Jan 1, 2025
- Research output includes QuaRot (outlier-free 4-bit LLM inference), Landmark Attention (landmark-attention repo), DenseFormer, and 'Masked Training of Neural Networks with Partial Gradients' (AISTATS 2022) - work in efficient LLM architectures and training.Apr 1, 2024
- IOI 2015 (Iran): silver medal, rank 63/322, absolute score 363.02 (60.50%).Jul 1, 2015
- Competitive-programming record beyond IOI: ICPC World Finals 2016 (rank 27) and 2018 (rank 31), Bayan Programming Contest bronze (2015), CodeChef SnackDown 12th (2016); $400 lifetime prize money. Codeforces handle 'matrix' (Master).Jan 1, 2018
Team Leader: Hossein Hadian - Topic: Persian Automatic Speech Recognition and Text to Speech ⋄ Improved ASR performance and implemented new features such as Diarization using Kaldi
Supervisor: Mahdieh Soleymani Worked on a novel neural network model in which the architecture is learned automatically during the training process.
Supervisor: Dan Alistarh ⋄ Worked on applying second-order optimizers (e.g., LBFGS) to train neural networks. ⋄ Designed an access-distribution adaptive data structure called the s
Competition record
Iran · IOI