Serre AI is an independent AI research lab in NYC that studies reasoning in AI systems through computational complexity, verification and large-scale empirical evaluation of language models.
Our research platform, Serre, is in closed testing. Request access
Our research is organised in three standing programs. Each pursues its own questions; all three are held to the same standard of evidence.
Which reasoning problems are hard for language models, and why?
We ask whether the systematic failures of model reasoning follow from the computational complexity of the problems themselves, rather than from scale or training, and what complexity theory predicts about where reasoning should break down.
When can an answer be checked more cheaply than it can be produced?
We study what can be verified, at what cost, and what a reliable check makes possible for systems that reason.
Does the theory hold when it is tested on real models?
We design evaluations of language models that isolate one reasoning question at a time, and size them in advance so that the result can confirm or refute what the theory predicts.
Rigour is how we work. Theory says what should happen; evidence decides whether it does.
A claim we make points to what it rests on: a proof, a measurement or a quoted passage. Anyone reading it can check it, and so can we.
We favour theory that says in advance where reasoning should succeed and where it should fail, so that an experiment can prove it wrong.
Predictions are tested across enough models, tasks and instances to tell a real effect from noise.
Serre is the research platform we build and run our own research on. It is in closed testing.
Gathers a field's papers through their citations, reads the core of the field in full, and records what each one claims with the exact passage it comes from.
Tests a research direction for novelty, feasibility, the soundness of its formal claims and its fit at real venues. Each judgment is a separate step with its evidence on file.
Places committed work at venues whose deadlines it can meet, with a fallback venue if a paper is rejected.
Logs every decision with the evidence it rested on and who made it, so it can be checked later.
Serre AI makes and stewards projects in the open, each with its own public repository, licence and governance.
Papers and preprints from our research programs.
Our first publications are in preparation.
Papers and preprints from our research programs will be listed here as they are released.
We welcome researchers interested in collaboration.
If your work touches the complexity of reasoning, verification or the evaluation of language models, we would like to hear from you. Write to us at