Mathias Lechner
I am Co-founder and CTO of Liquid AI, a startup founded in 2023 to focus on building efficient foundation models at every scale. We raised a Series A led by AMD, Shopify, Samsung, and G42 at a valuation of $2.3B. We build and release state-of-the-art language models called LFM2 at 1B parameters and below that run efficiently on CPU and edge devices. LFM2 outperform similar sized models such as Llama3 and Qwen3 at both speed and performance.
I am also a Research Affiliate at the Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT working with Prof. Daniela Rus. I completed my PhD (2022) at the Institute of Science and Technology Austria (ISTA) supervised by Tom Henzinger. Prior to my PhD, I received my master (2017) and bachelor (2016) degrees in Computer Science at the Vienna University of Technology (TU Wien).
Together with my colleague Djordje Zikelic I received the 2023 Outstanding Scientific Achievement Award at the Institute of Science and Technology Austria for our work on proving safety in stochastic machine learning systems. With my colleague Ramin Hasani, I got the Hyperion Research 2022 HPC Innovation Excellence Award for my work on Liquid neural and was honored as Outstanding Reviewer at ICRA 2021. I co-led the team that won the F1-Tenth - Autonomous Racing Grand Prix at IFAC 2020, and I received the Distinguished Young Alumni award at TU Vienna in 2018 for my MS thesis.
Contact: mlechner (at) mit (dot) edu
Last update: July 11, 2025
Publications
-
Continuous Autoregressive Generation with Mixture of Gaussians
ICML 2025 Workshop ES-FoMo-III, 2025
-
NeuralStoc: A Tool for Neural Stochastic Control and Verification
International Conference on Computer Aided Verification (CAV) - Tool Paper, 2025
-
SafeDiffuser: Safe Planning with Diffusion Probabilistic Models
International Conference on Learning Representations (ICLR) - Poster, 2025
-
Holistic Surgical Phase Recognition with Hierarchical Input Dependent State Space Models
arXiv preprint arXiv:2506.21330, 2025
-
Competitive Multi-Team Behavior in Dynamic Flight Scenarios
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024
-
Large scale dataset distillation with domain shift
International Conference on Machine Learning (ICML), 2024
-
Overparametrization helps offline-to-online generalization of closed-loop control from pixels
IEEE International Conference on Robotics and Automation (ICRA), 2774-2782, 2024
-
Learning with Chemical versus Electrical Synapses Does it Make a Difference?
IEEE International Conference on Robotics and Automation (ICRA), 15106-..., 2024
-
State-free inference of state-space models: The transfer function approach
arXiv preprint arXiv:2405.06147, 2024
-
Leveraging Low-Rank and Sparse Recurrent Connectivity for Robust Closed-Loop Control
International Conference on Learning Representations (ICLR), 2024
-
Gigastep - One Billion Steps per Second Multi-agent Reinforcement LearningConference on Neural Information Processing Systems (NeurIPS) - Dataset and Benchmark Track, 2023
-
Compositional Policy Learning in Stochastic Control Systems with Formal GuaranteesConference on Neural Information Processing Systems (NeurIPS), 2023
-
On the Size and Approximation Error of Distilled DatasetsConference on Neural Information Processing Systems (NeurIPS), 2023
-
Robust flight navigation out of distribution with liquid neural networksScience Robotics, 2023
-
Dataset Distillation with Convexified Implicit GradientsInternational Conference on Machine Learning (ICML), 2023
-
On the Forward Invariance of Neural ODEsInternational Conference on Machine Learning (ICML), 2023
-
A Learner-Verifier Framework for Neural Network Controllers and Certificates of Stochastic SystemsInternational Conference on Tools and Algorithms for the Construction and Analysis of Systems (TACAS), 2023
-
Liquid Structural State-Space ModelsInternational Conference on Learning Representations (ICLR), 2023
-
Infrastructure-based End-to-End Learning and Prevention of Driver FailureIEEE International Conference on Robotics and Automation (ICRA), 2023
-
Revisiting the Adversarial Robustness-Accuracy Tradeoff in Robot LearningIEEE Robotics and Automation Letters (RA-L), 2023
-
Quantization-aware Interval Bound Propagation for Training Certifiably Robust Quantized Neural NetworksAAAI Conference on Artificial Intelligence (AAAI), 2023
-
Learning Control Policies for Stochastic Systems with Reach-avoid GuaranteesAAAI Conference on Artificial Intelligence (AAAI), 2023
-
Closed-form continuous-time neural networksNature Machine Intelligence, 2022
-
PyHopper-A Plug-and-Play Hyperparameter Optimization EngineHas it Trained Yet? NeurIPS 2022 Workshop, 2022
-
Are All Vision Models Created Equal? A Study of the Open-Loop to Closed-Loop Causality GapNeurIPS 2022 Machine Learning for Autonomous Driving Workshop (ML4AD), 2022
-
Mixed-Memory RNNs for Learning Long-term Dependencies in Irregularly-sampled Time SeriesMemory in Artificial and Real Intelligence (MemARI) NeurIPS 2022 Workshop, 2022
-
Infinite Time Horizon Safety of Bayesian Neural NetworksConference on Neural Information Processing Systems (NeurIPS), 2021
-
Causal Navigation by Continuous-time Neural NetworksConference on Neural Information Processing Systems (NeurIPS), 2021
-
On-Off Center-Surround Receptive Fields for Accurate and Robust ImageInternational Conference on Machine Learning (ICML), 2021
-
Adversarial Training is Not Ready for Robot LearningIEEE International Conference on Robotics and Automation (ICRA), 2021
-
Scalable Verification of Quantized Neural NetworksAAAI Conference on Artificial Intelligence (AAAI), 2021
-
Liquid Time-constant NetworksAAAI Conference on Artificial Intelligence (AAAI), 2021
-
On the Verification of Neural ODEs with Stochastic GuaranteesAAAI Conference on Artificial Intelligence (AAAI), 2021
-
Neural circuit policies enabling auditable autonomyNature Machine Intelligence, 2020
-
The Natural Lottery Ticket Winner: Reinforcement Learning with Ordinary Neural CircuitsInternational Conference on Machine Learning (ICML), 2020
-
Learning Representations for binary classification without backpropagationInternational Conference on Learning Representations (ICLR), 2020
-
An SMT Theory of Fixed-Point ArithmeticInternational Joint Conference on Automated Reasoning (IJCAR), 2020
-
Gershgorin Loss Stabilizes the Recurrent Neural Network Compartment of an End-To-End Robot Learning SchemeIEEE International Conference on Robotics and Automation (ICRA), 2020
-
How Many Bits Does it Take to Quantize Your Neural Network?International Conference on Tools and Algorithms for the Construction and Analysis of Systems (TACAS), 2020
-
Lagrangian Reachtubes: The Next GenerationIEEE Conference on Decision and Control (CDC), 2020
-
Response characterization for auditing cell dynamics in long short-term memory networksInternational Joint Conference on Neural Networks (IJCNN), 2019
-
Designing worm-inspired neural networks for interpretable robotic controlIEEE International Conference on Robotics and Automation (ICRA), 2019