Sahand Sharifzadeh

Staff Research Scientist · Gemini · Google DeepMind, London

Sahand Sharifzadeh

Building frontier multimodal and reasoning models across 12+ years of research in foundation models, computer vision, deep RL, and robotics.

I developed a key recipe used in Gemini 4 Argon and contributed to the Gemini 2 and 3 model series and Flamingo. Across the Gemini Mid-Training and Thinking areas, I drove research projects on reinforcement learning efficiency and scaling, inference-time scaling, quantization in post-training, and strategic reasoning in games, and co-led research on self-improving VLMs.

Earlier, my research spanned looped graph transformers for iterative scene understanding (Classification by Attention, 2020), knowledge graph embeddings (PyKEEN), and deep inverse reinforcement learning for autonomous driving.

Google Scholar
Portrait of Sahand Sharifzadeh
17,000+
citations on Google Scholar
12+ yrs
foundation models, vision, RL & robotics
Gemini 2–4
mid-training & thinking
Ph.D.
summa cum laude, LMU Munich

Recently

  • Gemini 4 Argon is now available to everyone, after its announcement on 30 September. I was a core member of the Gemini Mid-Training team and developed a key recipe used in the model. Google blog ↗

  • Game Arena: strategic LLM evaluation in competitive environments (Kaggle). Technical report on arXiv. arXiv ↗

  • Gemini 2.5 technical report released. arXiv ↗

  • Synth²: self-improving VLMs with synthetic captions and image embeddings (co-lead). arXiv ↗

  • Ultra-high-granularity detector simulation published in Nature Communications. Paper ↗

01 · 2006 – present

Research trajectory

Twenty years, six threads: from building robots in Tehran to training Gemini in London. Each point is a project; the lines show how one led to the next.

Select a point to read about the project.

Read the trajectory as text

RL & post-training

  • 2014 – 2016, 2022 Inverse RL & Autonomous Driving. Motion planning with Inverse RL and Deep Q-Networks (NeurIPS 2016 Workshop, TUM CVG); Robotic pouring via curriculum RL (PourNet, IEEE/RSJ IROS 2022); 1st Place, ACM UIST 2013 Student Innovation Contest (Cool Print).
  • 2024 – 2025 Gemini Thinking (2 & 3 Series). Contributor across the Gemini 2 and 3 series; Drove research explorations on Thinking: inference-time scaling and post-training quantization.
  • 2025 – 2026 Gemini Games (Game Arena). Core contributor to Gemini Games: strategic reasoning and evaluation in games (Kaggle Game Arena).
  • 2025 – 2026 Gemini Mid-Training. Core member of the Gemini Mid-Training team at Google DeepMind; developed a key recipe used in Gemini 4 Argon.

Multimodal & vision–language

  • 2007 GooyaSaz (Persian OCR & TTS). 1st Place Khwarizmi Youth Innovation Award (Tehran, Dec 2007): Farsi Optical Character Recognition (OCR) and Text-to-Speech software converting scanned documents into speech for blind people.
  • 2019 – 2023 The Tensor Brain. The Tensor Brain: A Unified Theory of Perception, Memory, and Semantic Decoding (Neural Computation 2023 & CCN 2019).
  • 2018 – 2023 Scene Graphs & Relationformer. Ph.D. at LMU Munich, thesis graded Summa Cum Laude: First-author AAAI (2021, 2022) and ICPR (2021) papers on scene graph generation & depth maps; Relationformer (ECCV 2022); InstanceFormer video segmentation (AAAI 2023); Prior-RadGraphFormer (MICCAIW 2023).
  • 2021 – 2023 Flamingo VLM. Research Scientist Intern at DeepMind (Apr – Sep 2021): Flamingo visual language model for few-shot learning (NeurIPS 2022, 10,000+ citations); Do DALL-E and Flamingo Understand Each Other? (ICCV 2023).
  • 2024 Synth² (Self-Improving VLMs). Co-lead & first author at Google DeepMind: self-improving VLMs trained with synthetic captions and image embeddings.

Computer vision

  • 2014 – 2017 Stereo 3D Reconstruction. Variational stereo image 3D reconstruction for automatic determination of pruning mass in grapevine management and breeding (Australian Journal of Grape and Wine Research, 2017; TUM Computer Vision Group with Prof. Daniel Cremers).
  • 2015 – 2016 Medical Registration (MRFs). MRF-based image registration for intraoperative ultrasound brain-shift compensation towards robotic image-based neurosurgery (TUM CAMP, 2015–2016; 3rd HBP Workshop).

Knowledge graph embeddings

  • 2020 – 2021 Knowledge Graphs (PyKEEN). Co-author of: The PyKEEN knowledge graph embedding library (JMLR 2021); MemeGraphs (ICDAR 2023); Unsupervised KG-to-text generation (EMNLP 2020).
  • 2020 – 2021 Large-scale KGE Evaluation (TPAMI). Co-author of “Bringing Light Into the Dark”, a large-scale evaluation of knowledge graph embedding models under a unified framework (IEEE TPAMI 2021).

AI for Science

  • 2020, 2024 Genomics & Protein QA. MutLX platform for clone-specific genomic mutation discovery (eLife 2020); PQA: Zero-Shot Protein Question Answering for Free-Form Scientific Enquiry with Large Language Models (2024).
  • 2023 – 2024 Particle Detector Simulation. Ultra-High-Granularity Detector Simulation with Intra-Event Aware Generative Adversarial Network and Self-Supervised Relational Reasoning (Nature Communications 2024).

Robotics

  • 2006 Guard Robot (Safety Robotics). 1st Place Khwarizmi Youth Innovation Award (Tehran, Dec 2006): Design and build of an autonomous mobile robot for industrial safety applications.
  • 2011 – 2013 @Home Service Robot & Kinect HCI. Technical Team Lead of the Sharif @Home RoboCup Team: 1st Place AUTCup 2012, RobX & SUT TDP papers, gesture & speech recognition; Research Assistant at the NUS HCI Lab (2011): 3D Kinect hand gesture alignment.

02 · Career

Experience

  1. Aug 2022 – PresentLondon, United Kingdom

    Staff Research Scientist — Gemini Team Google DeepMind

    • Gemini Mid-Training: Core member of the Gemini Mid-Training team; developed a key recipe used in Gemini 4 Argon.
    • Gemini Thinking: Contributor across the Gemini 2 and 3 series; drove research explorations on Thinking (inference-time scaling and post-training quantization).
    • Gemini Games: Core contributor; strategic reasoning and evaluation in games (Kaggle Game Arena).
    • Synth²: Co-lead; self-improving VLMs trained with synthetic captions and image embeddings.
    Transformers — M2L Summer SchoolMediterranean Machine Learning Summer School (2023)arXivSynth²: Synthetic Captions & Image EmbeddingsKaggleGame Arena: Strategic LLM EvaluationarXivGemini 2.5 Technical Report
  2. Apr 2021 – Sep 2021London, United Kingdom

    Research Scientist Intern DeepMind

    Contributed to Flamingo, a visual language model for few-shot multimodal learning (NeurIPS 2022, 10,000+ citations).

    NeurIPSFlamingo: Visual Language Model for Few-Shot Learning
  3. 2018 – 2022Munich, Germany

    Doctoral Researcher LMU Munich

    Scene graph generation, Relationformer, The Tensor Brain, and co-author of the PyKEEN knowledge graph embedding library—with publications in Nature Communications, NeurIPS, ICCV, ECCV, AAAI, IEEE TPAMI, JMLR, eLife, EMNLP, and Neural Computation.

    Scene Graph Classification with Prior KnowledgeAdvanced School of Computing — ASoC (2021)JMLRPyKEEN 1.0: Knowledge Graph EmbeddingsNeural Comp.The Tensor Brain: Perception & Memory
  4. Oct 2013 – Apr 2017Munich, Germany

    Research and Teaching Assistant Technical University of Munich (TUM)

    • Autonomous Driving & Computer Vision (Computer Vision Group, Oct 2013 – Apr 2017): Motion planning with inverse RL and deep Q-networks; stereo 3D reconstruction for grapevine management; CUDA Programming TA (“GPU Programming for Computer Vision”, Sep 2014).
    • Medical Imaging (Computer Aided Medical Procedures Lab — CAMP, May 2015 – Mar 2016): MRF-based image registration for ultrasound brain-shift compensation in neurosurgery.
    Learning to Drive using Deep Inverse RLHuman Brain Project School of Future ComputingIntra-operative Ultrasound Brain-Shift CompensationHuman Brain Project Education Workshop — Manchester
  5. 2012 – 2013Tehran, Iran

    Technical Team Lead — @Home RoboCup Team Sharif University of Technology

    Design and build of a service robot; gesture and speech recognition.

  6. 2011Singapore

    Research Assistant — HCI Lab National University of Singapore (NUS)

    Hand gesture recognition from Kinect data using sequence alignment.

03 · Foundation

Education

  1. 2018 – 2022

    Ph.D. in Computer Science (Artificial Intelligence)

    Ludwig Maximilian University of Munich (LMU), Germany

    Thesis: Summa Cum Laude (0.7 — The Highest Distinction)

    Thesis: On the Importance of Symbol Grounding and Top-Down Processes in Computer Vision
    Supervisor: Prof. Volker Tresp

  2. 2013 – 2016

    M.Sc. in Computer Science

    Technical University of Munich (TUM), Germany

    Thesis: High Distinction (1.1)

    Major: Computer Vision, Robotics & AI
    Supervisors: Prof. Daniel Cremers, Prof. Rudolph Triebel

  3. 2008 – 2012

    B.Sc. in Computer Science

    Shahid Beheshti University, Iran

    Thesis: High Distinction (20/20)Top 3 of Class

    Thesis Supervisor: Prof. Shengdong Zhao (National University of Singapore)

04 · Recorded

Talks & lectures

Transformers Lecture — Sahand Sharifzadeh at M2L Summer School 2023 · Google DeepMind

Transformers — Mediterranean Machine Learning (M2L) Summer School

Invited summer school lecture covering Transformer architectures, attention mechanisms, and multimodal foundation models.

 ASoC 2021: Classification by Attention: Scene Graph Classification with Prior Knowledge — Sahand Sharifzadeh 2021 · LMU Munich

Classification by Attention: Scene Graph Classification with Prior Knowledge

Advanced School of Computing (ASoC 2021, IPM) — talk on symbol grounding, attention mechanisms, and prior-knowledge scene graph generation (AAAI 2021).

 Sharifzadeh Sahand - Learning to Drive using Deep Inverse Reinforcement Learning 2016 · TUM Computer Vision Group

Learning to Drive using Deep Inverse Reinforcement Learning

Human Brain Project School of Future Computing — presentation on deep inverse RL and Deep Q-Networks for autonomous driving.

 Sharifzadeh Sahand: Intra-operative ultrasound based brain-shift compensation 2016 · TUM CAMP Lab

Intra-operative Ultrasound-Based Brain-Shift Compensation Towards Robotic Neurosurgery

3rd Human Brain Project Education Workshop on Future Computing (Manchester) — MRF-based medical image registration.

05 · Papers

Publications

A curated list. The complete, up-to-date record is on Google Scholar.

  1. 2026Technical Report · arXivKaggle Game Arena ↗

    Game Arena: Strategic LLM Evaluation in Competitive Environments

    B Doerschuk-Tiberi, Y Yan, …, S Sharifzadeh, …, O Firat, M Chen

  2. 2025Technical Report

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Gemini Team (incl. S Sharifzadeh)

  3. 2024arXiv preprintFirst Author & Co-Lead

    Synth²: Boosting Visual-Language Models with Synthetic Captions and Image Embeddings

    S Sharifzadeh, C Kaplanis, S Pathak, D Kumaran, A Ilic, J Mitrovic, C Blundell, A Banino

  4. 2024Nature Communications

    Ultra-High-Granularity Detector Simulation with Intra-Event Aware Generative Adversarial Network and Self-Supervised Relational Reasoning

    B Hashemi, N Hartmann, S Sharifzadeh, J Kahn, T Kuhr

  5. 2024arXiv preprint

    PQA: Zero-Shot Protein Question Answering for Free-Form Scientific Enquiry with Large Language Models

    EM Carrami, S Sharifzadeh

  6. 2023AAAI

    InstanceFormer: An Online Video Instance Segmentation Framework

    R Koner, T Hannan, S Shit, S Sharifzadeh, M Schubert, T Seidl, V Tresp

  7. 2023ICCV

    Do DALL-E and Flamingo Understand Each Other?

    H Li, J Gu, R Koner, S Sharifzadeh, V Tresp

  8. 2023Neural Computation

    The Tensor Brain: A Unified Theory of Perception, Memory, and Semantic Decoding

    V Tresp, S Sharifzadeh, H Li, D Konopatzki, Y Ma

  9. 2023ICDAR

    MemeGraphs: Linking Memes to Knowledge Graphs

    V Kougia, S Fetzel, T Kirchmair, E Çano, SM Baharlou, S Sharifzadeh, B Roth

  10. 2023MICCAI Workshop (GRAIL)

    Prior-RadGraphFormer: A Prior-Knowledge-Enhanced Transformer for Generating Radiology Graphs from X-Rays

    Y Xiong, J Liu, K Zaripova, S Sharifzadeh, M Keicher, N Navab

  11. 2022NeurIPS10,000+ Citations

    Flamingo: A Visual Language Model for Few-Shot Learning

    JB Alayrac, J Donahue, P Luc, A Miech, …, S Sharifzadeh, et al.

  12. 2022ECCV

    Relationformer: A Unified Framework for Image-to-Graph Generation

    S Shit, R Koner, B Wittmann, J Paetzold, I Ezhov, H Li, J Pan, S Sharifzadeh, et al.

  13. 2022AAAIFirst Author

    Improving Scene Graph Classification by Exploiting Knowledge from Texts

    S Sharifzadeh, SM Baharlou, M Schmitt, H Schütze, V Tresp

  14. 2022IEEE/RSJ IROS

    PourNet: Robust Robotic Pouring Through Curriculum and Curiosity-Based Reinforcement Learning

    E Babaians, T Sharma, M Karimi, S Sharifzadeh, E Steinbach

  15. 2022Ph.D. Dissertation · LMU MunichSumma Cum Laude

    On the Importance of Symbol Grounding and Top-Down Processes in Computer Vision

    S Sharifzadeh

  16. 2021AAAIFirst Author

    Classification by Attention: Scene Graph Classification with Prior Knowledge

    S Sharifzadeh, SM Baharlou, V Tresp

  17. 2021ICPRFirst Author

    Improving Visual Relation Detection using Depth Maps

    S Sharifzadeh, SM Baharlou, M Berrendorf, R Koner, V Tresp

  18. 2021IEEE TPAMI

    Bringing Light Into the Dark: A Large-scale Evaluation of Knowledge Graph Embedding Models Under a Unified Framework

    M Ali, M Berrendorf, CT Hoyt, L Vermue, M Galkin, S Sharifzadeh, et al.

  19. 2021JMLR

    PyKEEN 1.0: A Python Library for Training and Evaluating Knowledge Graph Embeddings

    M Ali, M Berrendorf, CT Hoyt, L Vermue, S Sharifzadeh, et al.

  20. 2020EMNLP

    An Unsupervised Joint System for Text Generation from Knowledge Graphs and Semantic Parsing

    M Schmitt, S Sharifzadeh, V Tresp, H Schütze

  21. 2020eLife

    A Highly Accurate Platform for Clone-Specific Mutation Discovery Enables the Study of Active Mutational Processes

    M KaramiNejadRanjbar, S Sharifzadeh, et al.

  22. 2019CCN

    A Model for Perception and Memory

    V Tresp, S Sharifzadeh, D Konopatzki

  23. 2017Australian Journal of Grape and Wine ResearchStereo 3D Reconstruction

    Automatic Image-Based Determination of Pruning Mass as a Determinant for Yield Potential in Grapevine Management and Breeding

    A Kicherer, M Klodt, S Sharifzadeh, D Cremers, R Töpfer, K Herzog

  24. 2016NeurIPS Workshop on Deep Learning for Action and InteractionFirst Author

    Learning to Drive using Inverse Reinforcement Learning and Deep Q-Networks

    S Sharifzadeh, I Chiotellis, R Triebel, D Cremers

  25. 20163rd Human Brain Project Education Workshop on Future ComputingFirst Author

    Intra-operative Ultrasound-Based Brain-Shift Compensation Towards Robotic Image-Based Neurosurgery

    S Sharifzadeh, SA Ahmadi, F Milletari, N Navab

  26. 2012–2013AUTCup & Khwarizmi International Robotic Competitions (@Home League)

    RobX & SUT Service Delivery Robots League Team Description Papers

    A Keipour, S Sharifzadeh, E Babaians, et al.

06 · Recognition

Honors & awards

Languages Persian · English · Azerbaijani · German · Turkish