Hybrid and neuro-symbolic reasoning
Explore combinations of language-model inference, domain rules, constraints, knowledge representation and computational engines to improve controllability and inspectability.
Trustworthy Agentic AI
Researching better reasoning, grounding and coordination.
Our research agenda focuses on practical limits of agentic systems in consequential domains and ways to make them more structured, testable and transparent.
Multi-agent simulation and coordination
Study role-specialized agents, interactions, objective constraints and evaluation strategies for complex simulations.
Knowledge grounding
Investigate retrieval, graph-based context, provenance and access-aware use of unstructured and structured evidence.
Human-AI governance
Study confidence and abstention signals, escalation, overrides, review interfaces and evidence presentation.
Domain evaluation
Create realistic test conditions, adversarial edge cases and domain-expert review protocols before practical deployment.
A closer look
Rigorous inquiry.
Practical relevance.
01Neuro-symbolic and hybrid reasoning
Natural-language models are powerful, but some questions require explicit structure, constraints and computation.
- Explore knowledge graphs, ontologies, policies and machine-readable rules.
- Keep mathematical calculations in testable deterministic components.
- Measure whether hybrid methods improve traceability and consistency versus model-only baselines.
02Multi-agent collaboration
Different roles can improve task decomposition, but agent agreement is not itself evidence of correctness.
- Study coordination, debate, critique and retrieval-driven validation.
- Detect agent loops, unsupported assertions and conflicting interpretations.
- Require independent reference checks and expert adjudication for consequential conclusions.
03Evaluation and trustworthy deployment
Research should result in measurable knowledge about what works and what does not.
- Create benchmark datasets with explicit provenance, scope and permitted use.
- Evaluate grounding, reproducibility, uncertainty, fairness and task completion.
- Public research claims are supported by approved evidence.