Research Program
Noether Research develops theoretical and computational frameworks spanning AI safety, biosecurity, biological intelligence, and complex systems. The program advances evaluation reliability, interaction safety, scientific rigor, and long-term resilience.
Research on evaluation reliability, interaction safety, diagnostic frameworks, and scientific reasoning in frontier AI systems. This work studies how advanced systems behave under evaluation, how high-stakes human-AI interactions become unsafe, how safety-relevant failures can be identified before they become operationally consequential, and how AI systems participate in scientific inquiry, uncertainty tracking, and verification.
- Evaluation reliability: Methods for identifying gaps between evaluation settings, deployment contexts, and observed system behavior under uncertainty, pressure, or distributional shift.
- Interaction safety and recoverability: Analysis of conversational irreversibility, option value preservation, escalation, repair windows, and state transitions in high-stakes human-AI interaction.
- Failure-mode diagnostics: Systematic study of evaluation gaps, unsafe interaction dynamics, capability-relevant ambiguity, and breakdowns between intended and observed behavior.
- Scientific reliability and verification: Study of how AI systems track assumptions, represent uncertainty, recover from errors, and maintain epistemic integrity across long technical reasoning chains, including principles for preserving experimental accountability and independent verification in AI-assisted scientific workflows.
- Mathematical physics and model-assisted inquiry: Use of physics and applied mathematics as test cases for AI-assisted derivation, analogy formation, conjecture exploration, and proof-sketch reliability.
- Responsible disclosure: Frameworks for staged disclosure, capability-relevant reporting, and risk communication with trusted reviewers, funders, or relevant institutions.
Research on AI-enabled biosecurity, threat modeling, defensive workflows, and institutional risk surfaces. This work focuses on how advanced AI systems may alter biological-risk landscapes, where defensive systems may fail, and which near-term interventions could reduce risk while preserving beneficial scientific work.
- Biosecurity threat modeling: Structured analysis of how AI systems may change capability access, workflow risk, institutional exposure, and defensive bottlenecks across biological research and deployment contexts.
- Methods survey: Mapping current and emerging approaches for AI-enabled biological-risk assessment, detection, screening, attribution, and response.
- Defensive workflow design: Identifying practical interventions that improve safety in research review, synthesis screening, procurement, laboratory governance, and institutional decision processes.
- Screening and governance bottlenecks: Studying where existing screening systems, review processes, and operational safeguards may become insufficient as AI-enabled capability increases.
- Near-term intervention mapping: Prioritizing tractable first-phase projects that reduce biological-risk amplification while supporting responsible scientific progress.
Research on biological intelligence, information fidelity, repair, reconstruction, and alternative cognitive architectures. This axis studies biological systems as information-bearing, adaptive, and partially reconstructable systems, with applications to neuroscience, regeneration, aging, reconstruction fidelity, and advanced AI safety analogies.
- Biological information fidelity: Information-theoretic framing for what is preserved, degraded, reconstructed, or lost across biological repair, regeneration, aging, and state transition. Biological systems are treated as information-bearing structures subject to Shannon-type analysis across time.
- White-matter topology and neural architecture: Geometric methods applied to white-matter networks, diffusion MRI, and system-level organization in biological cognition and neurodegeneration.
- Alternative cognitive architectures: Comparative study of non-human intelligence, embodied sensing, biological cognition, and non-language-centric forms of agency and perception.
- Substrate-level safety analogies: Careful use of biological regulation, repair, immune response, network dynamics, and cognitive stability as reference systems for advanced AI safety and diagnostic frameworks.
Research on decision theory, crisis dynamics, irreversibility, and reliability in high-stakes complex systems. This axis connects mathematical and computational models of decision trajectories with real-world systems such as markets, governance processes, compute infrastructure, and safety-critical human-AI interaction.
- Decision dynamics: Modeling escalation, reversibility, commitment, uncertainty, and repair windows in complex decision trajectories.
- Irreversibility and option collapse: Study of how systems move from high-optionality states into constrained regimes where recovery becomes difficult, costly, or impossible.
- Electricity-market and infrastructure modeling: Application of electricity-market modeling and infrastructure-system experience to reliability, resource allocation, and systemic risk under accelerating technological demand.
- Compute governance foundations: Decision-system foundations for compute governance, priority access, infrastructure externalities, and institutional response under accelerating frontier AI deployment.
Some technical details, diagnostics, and related materials are shared only with qualified reviewers, funders, and collaborators. Access is handled through a structured review process to support responsible disclosure and research integrity. Review access details.