PeopleofSuperintelligence

Researcher

Ido Amos

Ido Amos is a machine-learning researcher whose paper 'Never Train from Scratch' won an Outstanding Paper award at ICLR 2024. In January 2022, he coauthored ASOCEM for automatic contamination segmentation in cryo-EM, and in 2023 at Tel Aviv University he studied overfitting in graph neural networks for a paper at ICML 2024. Posted on arXiv in October 2023 with Jonathan Berant and Ankit Gupta, 'Never Train from Scratch' showed that properly pretrained vanilla Transformers match S4 on Long Range Arena, and an extended abstract appeared in the Sister Conferences Best Papers track of IJCAI in 2025. In 2026, he coauthored papers on latent reasoning with Thinking States, while affiliated with Google Research and the Hebrew University of Jerusalem, and on latent state propagation with LIFT. He is affiliated with the Hebrew University of Jerusalem.

Badges

Timeline

  1. January 2022

    Segmenting contaminations in cryo-EM

    In January 2022, he coauthors 'ASOCEM: Automatic Segmentation Of Contaminations in cryo-EM' with Amitay Eldar and Yoel Shkolnisky. Their method automatically detects and segments contaminations in cryo-EM micrographs using only the approximated particle size as input.

  2. 2023–2024

    Overfitting in graph networks

    While at the School of Electrical Engineering of Tel Aviv University in 2023, he coauthors 'Graph Neural Networks Use Graphs When They Shouldn't' with Maya Bechler-Speicher, Ran Gilad-Bachrach and Amir Globerson. The study shows that graph neural networks tend to overfit the given graph structure, and it appears at ICML 2024, the 41st International Conference on Machine Learning.

  3. October 2023

    Never Train from Scratch

    In October 2023, he posts 'Never Train from Scratch: Fair Comparison of Long-Sequence Models Requires Data-Driven Priors' on arXiv with Jonathan Berant of Tel Aviv University and Ankit Gupta of IBM Research. Proper pretraining lets vanilla Transformers match S4 on Long Range Arena, and the work improves the best reported state space model result on PathX-256 by 20 absolute points.

  4. 2024

    An ICLR Outstanding Paper

    Presented as an oral at ICLR 2024 in Vienna in May 2024, 'Never Train from Scratch' receives an Outstanding Paper award. It is one of 5 award winners chosen by the conference's Outstanding Paper Committee, whose write-up notes that training transformer models from scratch leads to an under-estimation of their performance.

  5. 2025

    A best-paper abstract at IJCAI

    An extended abstract of 'Never Train from Scratch' appears in 2025 in the Sister Conferences Best Papers track of the Thirty-Fourth International Joint Conference on Artificial Intelligence, on pages 10846–10851. It presents the work as an ICLR 2024 Outstanding Paper Award paper.

  6. February 2026

    Thinking States for latent reasoning

    In February 2026, he coauthors 'Latent Reasoning with Supervised Thinking States' with Avi Caciularu, Mor Geva, Amir Globerson, Jonathan Herzig, Lior Shani and Idan Szpektor. Affiliated with Google Research and the Hebrew University of Jerusalem, he helps propose Thinking States, a method that generates thinking tokens while the input is being processed.

  7. As of October 2026

    Research at Hebrew University

    He is affiliated with the Hebrew University of Jerusalem. In September 2026, he coauthors 'Pretraining Latent Information Feedback Transformers with Teacher Supervision' with Dor Tirosh and Mor Geva, which introduces the LIFT architecture and training method that lets language models propagate state across generation.

Career

  1. NowHebrew University of Jerusalem
  2. From 2026Google Research
  3. From 2023Tel Aviv University

From the biography.

  1. 16% of AI career moves in the 2020s went from a university to a companyIncludes Ido Amos. We estimate that 16% of AI career moves in the 2020s went from a university to a company; 29% of moves starting at a university went to a company, and 76% ended the university role.
  2. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other universityIncludes Hebrew University of Jerusalem. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other university. We compare doctorate and 5+ year employment links, career timing and current workplaces for 187 major-prize and 1,111 best-paper winners.
  3. Google had the most people shaping AI every year from 2014 to 2025Includes Google. 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.
  4. Google has the highest combined h-index among AI organizationsIncludes Google. 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 Ido Amos yet. Send us a free-licensed photo.

  1. [2201.06978] ASOCEM: Automatic Segmentation Of Contaminations in cryo-EM arxiv.org. Checked 22 October 2026.
  2. [2309.04332] Graph Neural Networks Use Graphs When They Shouldn't arxiv.org. Checked 22 October 2026.
  3. Graph Neural Networks Use Graphs When They Shouldn't (arXiv HTML) arxiv.org. Checked 22 October 2026.
  4. Graph Neural Networks Use Graphs When They Shouldn't | Proceedings of Machine Learning Research proceedings.mlr.press. Checked 22 October 2026.
  5. [2310.02980] Never Train from Scratch: Fair Comparison of Long-Sequence Models Requires Data-Driven Priors arxiv.org. Checked 22 October 2026.
  6. Never Train from Scratch: Fair Comparison of long-sequence models requires data-driven priors (arXiv HTML) arxiv.org. Checked 22 October 2026.
  7. ICLR 2024 Awards iclr.cc. Checked 22 October 2026.
  8. ICLR 2024 Outstanding Paper Awards – ICLR Blog blog.iclr.cc. Checked 22 October 2026.
  9. Congratulations to the #ICLR2024 test of time and outstanding paper award winners - AIhub aihub.org. Checked 22 October 2026.
  10. Never Train from Scratch: Fair Comparison of Long-Sequence Models Requires Data-Driven Priors (Extended Abstract) | IJCAI ijcai.org. Checked 22 October 2026.
  11. Latent Reasoning with Supervised Thinking States (arXiv HTML) arxiv.org. Checked 22 October 2026.
  12. Pretraining Latent Information Feedback Transformers with Teacher Supervision (arXiv HTML) arxiv.org. Checked 22 October 2026.