What we work on
Mechanism claims are everywhere in machine learning, and most are asserted rather than tested — a circuit is said to implement a task, a feature is said to represent a variable. Mechanistic Research builds the methods that turn these claims into something you can measure and, if they’re wrong, falsify. The core idea is to bring the rigor of causal inference, philosophy of science, and measurement theory into mechanistic interpretability.
How we approach it
- Decompose the claim into the causal steps it actually depends on.
- Ask what each step requires to be true, and find the ground truth that would settle it.
- Test it directly — against real data, with a simple baseline the method has to beat and predictions registered before looking at outcomes.
We hold our own methods to this same standard.
Who we are
Mechanistic Research is an independent research group, not affiliated with any single institution. Members hold their own affiliations and collaborate under a shared program. It grew out of one researcher’s work across mechanistic interpretability and causal inference, and is opening up into a small collective working on the same questions.
We collaborate with labs and independent researchers, and we’re open to funded work aimed at better methods for validating claims about neural networks.
People
Elliot Tower
Primary interest is integrating causal inference and philosophy of science into applied domains: mechanistic interpretability, neuroscience, and clinical epidemiology.
Background:
- Machine Learning Engineer, Eluve Inc (2024–2026)
- Machine Learning Engineer, Swarm Labs (2023–2024)
- Maintainer / project manager, PettingZoo at Farama Foundation (2023–2024)
- Research intern, IESL, UMass (2021) · co-author on ICML 2022
- M.S. Computer Science, UMass Amherst (2020–2022)
- B.S. Mathematics & Philosophy, UMass Amherst (2016–2020)
Collaborators
Researchers contributing to specific projects are listed on those projects. If you’ve worked with us, your affiliation and links go here.
Getting involved
We’re a collective, not a closed lab. If you’re working on validating mechanism claims — in interpretability, neuroscience, causal inference, or applied methods — and want to collaborate, we’d like to hear from you. Contributors can join under the group’s affiliation on shared papers while keeping their own institutional identity.