Martin Weiss

Martin Weiss

Tiptree Systems · Mila - Quebec Artificial Intelligence Institute
6 previous affiliations
  • Tiptree SystemsSince 2023
    • Co-Founder · Since 2023
    • Postdoc · 2025–2027
    • PHD Candidate · 2021–2025
    • MSc Student · 2018–2020
    • Research Intern · 2018
  • The Computer Vision Foundation2020–2021
    • Tech Chair for the International Conference on Computer Vision (2021) · 2020–2021
    • Scientist-in-Residence · 2021
  • YesGraph2016–2017
    • Software Engineer · 2016–2017
    • Bachelor of Science (B.S.), Software Engineering · 2012–2015
27Machine learning papers tracked by Lacuna
  • Founding author
Is this you?

Research Overview

Martin Weiss focuses on artificial intelligence systems for scientific discovery, multi-agent modeling, and representation learning. His early work explored graph neural networks for biological data and mobile computing for accessibility. Recently, his research has shifted toward developing computational tools and large language model agents that accelerate literature synthesis, problem formulation, and the generation of novel research directions.

Research Trajectory

Weiss's early work addressed mobile computing for accessibility, beginning with A Survey of Mobile Computing for the Visually Impaired, which outlined how machine learning could support visually impaired individuals. This work touched on the research direction of Efficient Single Stage Object Detection for Edge Hardware and the broader area of Decentralized Model Training via Federated Learning. He subsequently developed simulated environments to train navigation systems in Navigation Agents for the Visually Impaired: A Sidewalk Simulator and Experiments.

Concurrently, Weiss investigated deep learning applications for genomics. He contributed to Towards Gene Expression Convolutions using Gene Interaction Graphs, which utilized protein-protein interactions to structure graph convolutional networks, advancing the area of Graph Neural Networks with Biological Structural Priors and Hierarchical Graph Pooling for Multiscale Representation Learning. A follow-up collaborative study, Is graph biased feature selection of genes better than random?, challenged prevailing assumptions in the field of Learning from Limited Data in Specialized Medical Domains and the area of The Redundancy of Topology in Graph Learning Benchmarks by demonstrating that random gene selection often performs as well as curated interaction graphs for predicting expression levels.

In 2021, Weiss participated in public health modeling efforts, joining the team behind Predicting Infectiousness for Proactive Contact Tracing, which evaluated digital contact tracing methods during the COVID-19 pandemic. His focus then expanded into multi-agent systems and representation learning. He collaborated on Latent Variable Sequential Set Transformers for Joint Multi-Agent Motion Prediction, proposing an encoder-decoder architecture for scene-consistent trajectory forecasting. He also introduced novel modular architectures in Neural Attentive Circuits, designing a general-purpose neural network with learned, sparse connectivity to improve out-of-distribution robustness. Extending his work in representation learning to 3D environments, he developed Visual question answering from another perspective: CLEVR mental rotation tests, a benchmark requiring models to perform mental rotation using 3D feature volumes.

More recently, Weiss's research has centered on large language models (LLMs) and their application to complex analytical tasks and scientific discovery. He explored information economics in Redesigning Information Markets in the Era of Language Models, simulating a marketplace where LLM agents inspect information prior to purchase. In the geospatial domain, he developed MapBot: A Multi-Modal Agent for Geospatial Analysis, an interactive system that orchestrates vision models for natural language querying of satellite imagery. He also contributed to SelvaBox: A high‑resolution dataset for tropical tree crown detection, advancing multi-resolution training for remote sensing.

His latest work focuses on computational methods to accelerate the research process itself. He co-developed DeLLMphi: A Multi-Turn Method for Multi-Agent Forecasting, which replaces human expert panels with LLMs to streamline iterative prediction. To address the challenge of generating novel scientific ideas, he co-authored Alien Science: Sampling Coherent but Cognitively Unavailable Research Directions from Idea Atoms, introducing a pipeline that decomposes literature into conceptual units to propose non-obvious research directions. He further formalized this approach by introducing Lacuna: A Research Map for Machine Learning Problem Formulation, a navigable knowledge graph of over 730,000 papers designed to help AI agents formulate high-quality research questions efficiently.

Current Directions

The most recent papers in the evidence indicate a strong focus on AI-driven scientific discovery and multi-agent systems. Recent work investigates how large language models can be orchestrated to automate complex, multi-step workflows, such as geospatial analysis and structured forecasting. A major emerging theme is the computational modeling of the research process itself. By structuring scholarly literature into granular concepts and navigable knowledge graphs, these systems aim to help AI agents and human researchers identify non-obvious connections, formulate concrete research proposals, and accelerate the synthesis of new knowledge.

Profile updated

See something wrong? Report a mistake