Research
That is the question my research program is organized around. When a model produces a chain of “thought,” is that reasoning faithful to what actually drove the answer, or is it a performance? Can it be bent by an adversary, and can it bend us? Can its conclusions be challenged and defended, the way we demand of human reasoning in any serious institution?
I evaluate AI reasoning along four properties, at every level from single model to agent to multi-agent system, using adversarial probes as the primary instrument.
Legibility & faithfulness — can you inspect it?
Reasoning should be inspectable, not opaque. A model's stated reasoning should be faithful to its actual inference process. I approach this with structured and vector-symbolic representations (see VSAX), neuro-symbolic argumentation (ACS 2022), and agentic verification: decomposing LLM outputs into claims, evidence, assumptions, and uncertainty.
Manipulation robustness — does it survive attack?
Can an adversary bend the model's reasoning, and is the model bending its user's? This is cognitive security (CogSLLaM, AAAI-SS 2026), and it's where my legal training does the most work. I explore questions of persuasion, hallucinations, overgeneralizations, jailbreaks and subtle behavioral nudges that influence human-AI interaction.
Contestability — can you challenge it?
A trustworthy conclusion is one that can be argued with: challenged, defended, and revised. I build computational argumentation systems (IntelliProof, AAAI 2026; ArgLib) and benchmarks that test whether models can genuinely engage with arguments rather than merely produce them. Evidence suggests that AI models are capable arguers and generating arguments for both sides unlocks reasoning a model otherwise can't access (EMNLP-W 2024).
Social reasoning — does it model other minds?
Systems that act among people must reason about what others believe, intend, and expect, including norms and conversational timing (EMNLP 2024; AAAI 2026). I explore whether AI systems maintain a Theory-of-mind, i.e., whether they can reason about others' beliefs, and whether these models can reason about implicit socio-cultural norms that we humans take for granted and incorporate into our own reasoning.
From models to agents to institutions
Each of these properties gets harder as systems move from answering questions to taking actions in the real world. My earlier work in cognitive robotics and reinforcement learning studied reliable action under uncertainty and novelty (SPOTTER, AAMAS 2021; RAPid-Learn, ICDL 2022); my current work carries those questions to LLM agents — uncertainty quantification for tool-using systems (AAMAS 2025) and verification of agentic reasoning. The direction I find most compelling is multi-agent architectures: a system of AI agents is an institution, with delegation, trust boundaries, and conflicts of interest. I explore how these systems can be regulated and governed.
The method: adversarial evaluation
The instrument behind all of this is the probe. Broadly, a probe is a targeted manipulation or measurement that forces a system to reveal what it won't volunteer. I believe interventions beat observations; a hostile test beats a friendly one. To me, uncertainty quantification and mechanistic interpretability are probe families, and dataset benchmarks are probes at scale.