SemEval 2026
2nd place finalist in political evasion detection
AsymVerify ranked second on Task 6 with an asymmetric confidence-gated verification pipeline.
85% Macro F1
Research in learning, language, and computation
Kaons
Papers and figures across reasoning, operator learning, scientific machine learning, connectomics, and evaluation.
kaons.com/April 2026/contact@kaons.org
Abstract
Kaons is a research archive for work on language-model reliability, operator learning theory, scientific machine learning, connectomics, and evaluation. It emphasizes papers, figures, and technical results over commentary.
Contents
- 1
Research archive
→Paper pages, figures, and the current research list.
- 2
Innovations
→Key contributions and standout results across the current papers.
- 3
Selected work
→A compact front-page survey across reasoning and scientific ML.
- A
Researcher access
→Private login for researchers and collaborators.
- B
OpenReview
→Public discussion and workshop paper record.
- C
Contact
→General contact and research correspondence.
Pioneer / Innovations
A few of the sharper claims and results that distinguish the current work.
current highlights
Operator Learning
First four-term finite-sample OOD transfer bound for FNOs
The FNO paper separates complexity, discretization, distribution shift, and mesh aliasing in one transfer result.
Four-term OOD decomposition
Reasoning Reliability
Residual drift, not contradiction, dominates multi-turn failure
The reasoning paper shows that after repair, most remaining errors come from satisfiable drift rather than outright inconsistency.
68–95% satisfiable drift
Connectomics
Species-scale comparative signatures from interpretable neuron features
The connectomics work recovers cross-species structure while exposing reconstruction artifacts introduced by missing-data handling.
176,914 neurons · 5 species
Selected work
A compact front-page survey. The full archive, paper pages, and figures live under the research section.
[01]
ICLR 2026 Workshop on Reasoning and Planning for LLMs
Drift Dominates Contradiction in Multi-Turn Constraint Reasoning
Shows that after solver-guided repair, the dominant failure mode is not contradiction but answers that violate a still-satisfiable maintained state.
816 problems · 4 open-weight models · 68–95% of residual errors are satisfiable drift
[02]
MathAI 2026
Four-Term Finite-Sample OOD Transfer Bound for Fourier Neural Operators
Derives a finite-sample transfer bound for nonlinear Fourier Neural Operators that separates complexity, discretization, distribution shift, and mesh aliasing.
Four channels · Darcy, Helmholtz, and Burgers diagnostics · aliasing made explicit
[03]
IEEE ICBCB 2026
Scalable Comparative Connectomics: Interpretable Machine Learning Reveals Evolutionary Signatures and Reconstruction Artifacts
Uses minimal neuron-level features to recover species-level signatures while showing how missing-data handling can create reconstruction artifacts.
176,914 neurons · 5 species · 92.2% Random Forest accuracy
[04]
SemEval 2026 Task 6
SemEval highlightAsymVerify at SemEval-2026 Task 6: Asymmetric Confidence-Gated Verification for Political Evasion Detection
A confidence-gated verification pipeline for political evasion detection where opposing verification passes both route corrections through the Ambivalent class.
SemEval 2026
2ndplaceTask 6 result
85% Macro F1 · rank 2 · +5.2 to +21.7 over zero-shot