Research on the reliability of learning systems.

Kaons does frontier research at the intersection of applied mathematics and machine learning, on the reliability of learning systems: how they fail, and where they can be trusted.

Plate. The decay K0 → π+π−, after a bubble-chamber photograph. The neutral kaon leaves no track; it is seen only where it decays.

About Kaons

Abstract
The work turns clean mathematical questions into concrete results: new generalization bounds in operator learning, a second-place system in a shared-task evaluation, and analyses that isolate how models actually fail.
Research areas

Applied mathematicsReasoningOperator learningScientific MLConnectomicsEvaluation

Founder
Founded in 2024 by Sebastien Kawada.
Correspondence
contact@kaons.com
Researcher accessGitHub
Four-panel comic explaining AsymVerify. Panel 1: a single pass gives a decent guess (55.9 Macro F1) but many hard cases stay unresolved. Panel 2: spend compute where it matters by sending only the uncertain cases for extra checking. Panel 3: asymmetric verification checks confident predictions for a downgrade and ambivalent ones for an upgrade, targeting the Ambivalent boundary. Panel 4: the result is higher accuracy and lower average cost, 0.85 Macro F1 and second place at SemEval.
Figure 1. AsymVerify, a confidence-gated verification pipeline: extra compute is spent only on the cases the model is uncertain about.SemEval 2026 · Task 6 · 0.85 Macro F1 · 2nd of 41 systems