# Mattia Rigotti

> Mattia Rigotti is a Senior Research Scientist at IBM Research in Zurich, working at the intersection of machine learning and neuroscience. Research areas: reasoning and alignment in large language models, neural operators for computational fluid dynamics, visual foundation models, trustworthy AI, and the geometry of neural representations in deep learning architectures and the brain. Previously Research Staff Member at the IBM T.J. Watson Research Center (2014–2020) and postdoctoral researcher in theoretical neuroscience at New York University and Columbia University. Ph.D. in Theoretical Neuroscience from Columbia University (2010), M.Sc. in Theoretical Physics from ETH Zurich.

## Career

- 2020–present: Senior Research Scientist, IBM Research – Zurich, Switzerland.
  - Technical lead for algorithmic innovation in IBM's neural operators for computational fluid dynamics effort (since 2024); GIST (NeurIPS 2026).
  - Initiated IBM's visual prompting research and served as its technical lead; shipped as VP Lab in IBM Maximo Visual Inspection (generally available 2025).
  - LLM alignment and reasoning: Alignment via Optimal Transport (NeurIPS 2024), Cognitive Tools (NeurIPS 2025).
  - Principal Investigator for IBM Research on EU Horizon grants (2020–present).
  - Supervises PhD students (TU Wien, ETH Zurich, EPFL) and master's interns.
- 2014–2020: Research Staff Member, IBM T.J. Watson Research Center, Yorktown Heights, New York.
  - Led the challenge team on neural network training under hardware and biological constraints.
  - Lead developer of the approximate nearest-neighbor retrieval pipeline for IBM MolFormer.
  - 1st place, CVPR 2020 VizWiz Grand Challenge (image captioning for blind users).
- 2013–2014: Associate Research Fellow, Columbia University.
- 2010–2013: Postdoctoral Fellow, Center for Neural Science, New York University.

## Education

- Ph.D. in Theoretical Neuroscience, Columbia University, 2010.
- M.Phil. in Neurobiology and Behavior, Columbia University, 2009.
- DEA in Theoretical Physics, Université Pierre et Marie Curie / École Normale Supérieure, Paris, 2004.
- M.Sc. in Theoretical Physics, ETH Zurich, 2005.

## Skills

LLM post-training, preference optimization, reinforcement learning, LLM reasoning, test-time scaling, vision-language models, visual prompting, foundation models, neural operators, scientific machine learning, graph neural networks, interpretability, model quantization, trustworthy AI, computational neuroscience, machine learning, deep learning, large language models, computational fluid dynamics

## Highlights

- IEEE CASS 2025 Best Paper Award: [Auditing and Generating Synthetic Data with Controllable Trust Trade-offs](https://doi.org/10.1109/JETCAS.2024.3477976), IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 2024.
- ACL 2023 oral presentation: [DARE: Towards Robust Text Explanations in Biomedical and Healthcare Applications](https://virtual2023.aclweb.org/paper_P1265.html), The 61st Annual Meeting of the Association for Computational Linguistics, 2023.
- NeurIPS 2022 oral presentation: [Compositional generalization through abstract representations in human and artificial neural networks](https://openreview.net/forum?id=177GzUAds8U), Advances in Neural Information Processing Systems 35, 2022.
- ECCV 2022 Workshop on Computer Vision for Civil and Infrastructure Engineering Best Paper Award: [Model-Assisted Labeling via Explainability for Visual Inspection of Civil Infrastructures](https://doi.org/10.48550/arXiv.2209.11159), European Conference on Computer Vision, 2022.
- [The importance of mixed selectivity in complex cognitive tasks](http://dx.doi.org/10.1038/nature12160), Nature, 2013.

## Key publications

- [SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs](https://icml.cc/virtual/2026/poster/62906). N. Avogaro, N. Debnath, L. Mi, T. Frick, J. Wang, Z. He, H. Hua, K. Schindler, M. Rigotti. International Conference on Machine Learning, ICML 2026. [PDF](https://www.matrig.net/publications/articles/avogaro2026.pdf)
  For vision-language models (VLMs), unstructured chain-of-thought entangles perception and reasoning, so small perceptual errors derail test-time scaling. SPARC decouples the two in a brain-inspired pipeline (explicit visual search to localize the relevant image regions, then reasoning conditioned on them), making inference-time scaling reliable without reinforcement learning or hand-crafted rewards.
- [Eliciting Reasoning in Language Models with Cognitive Tools](https://neurips.cc/virtual/2025/poster/119804). B. Ebouky, A. Bartezzaghi, M. Rigotti. Advances in Neural Information Processing Systems 38, NeurIPS 2025. [PDF](https://www.matrig.net/publications/articles/ebouky2025.pdf)
  Drawing on cognitive psychology and cognitive architectures, this work equips off-the-shelf large language models with a small set of modular 'cognitive tools' (understanding, recalling, examining, verifying) that the model calls to structure its reasoning. Orchestrating these tools substantially improves mathematical reasoning benchmarks, a transparent, training-free alternative to chain-of-thought elicited through reinforcement learning.
- [Auditing and Generating Synthetic Data with Controllable Trust Trade-offs](https://doi.org/10.1109/JETCAS.2024.3477976). B. Belgodere, P. Dognin, A. Ivankay, I. Melnyk, Y. Mroueh, A. Mojsilovic, J. Navratil, A. Nitsure, I. Padhi, M. Rigotti, J. Ross, Y. Schiff, R. Vedpathak, R.A. Young. IEEE Journal on Emerging and Selected Topics in Circuits and Systems, JETCAS (IEEE CASS 2025 Best Paper Award) 2024. [PDF](https://www.matrig.net/publications/articles/belgodere2024.pdf)
  Synthetic data from generative models promises privacy and fairness, but its trustworthiness is hard to verify. This work introduces a holistic auditing framework that quantifies synthetic datasets and their generators along fidelity, utility, privacy, and fairness, and uses those audits to steer generation toward controllable, application-specific trade-offs in trustworthy AI.
- [Compositional generalization through abstract representations in human and artificial neural networks](https://openreview.net/forum?id=177GzUAds8U). T. Ito, T. Klinger, D.H. Schultz, J.D. Murray, M.W. Cole, M. Rigotti. Advances in Neural Information Processing Systems 35, NeurIPS (Awarded an oral presentation) 2022. [PDF](https://www.matrig.net/publications/articles/ito2022.pdf)
  Combining whole-cortex human fMRI with artificial neural network modeling on a compositional task, this study pinpoints abstract, compositional representations as the shared mechanism behind generalization to novel tasks. Pretraining networks on task primitives induces the same low-dimensional geometry and yields human-like compositional (systematic) generalization.
- [The importance of mixed selectivity in complex cognitive tasks](http://dx.doi.org/10.1038/nature12160). M. Rigotti, O. Barak, M.R. Warden, X.-J. Wang, N.D. Daw, E.K. Miller, S. Fusi. Nature, 497:585-590, 2013. [PDF](https://www.matrig.net/publications/articles/rigotti2013.pdf)
  Prefrontal cortex neurons respond to seemingly arbitrary mixtures of task variables. Analyzing population recordings from monkeys in a memory task, this work showed that such nonlinear mixed selectivity is a feature, not noise: it raises the dimensionality of the neural representation, letting simple linear readouts implement a far richer repertoire of input-output functions than pure-selectivity neurons could. High-dimensional mixed selectivity thus offers a population-coding substrate for cognitive flexibility, linking single-neuron physiology to the geometry of neural population representations.

## Full record

- [Full publication list](https://www.matrig.net/llms-full.txt): all articles, conference abstracts, talks, and patents
- [Homepage](https://www.matrig.net/)
- [Google Scholar](https://scholar.google.com/citations?user=TmHt7CwAAAAJ)
- [ORCID](https://orcid.org/0000-0001-6466-2810)
- [Semantic Scholar](https://www.semanticscholar.org/author/2535094)
- [Wikidata](https://www.wikidata.org/wiki/Q57066174)
- [GitHub](https://github.com/matrig)
- [LinkedIn](https://www.linkedin.com/in/mattia-rigotti)
- [Mastodon](https://mastodon.social/@matrig)
- [Bluesky](https://bsky.app/profile/matrig.net)