Research
My research focuses on three questions:
AI Alignment
I use alignment in a broad sense: how can we develop models and agents that gain useful capabilities and generalise beyond familiar tasks and environments? My work here spans model adaptation, reasoning, embodied action, and agents that combine evidence and tools to solve complex tasks. Safety can also be part of what a system learns, rather than only an external safeguard.
- ELAN4D — generalisation in vision-language-action models; CoRL 2026.
- The Path Matters — a token-commitment policy learned from agreement with final decoded tokens, improving the speed–quality trade-off in diffusion language models; preprint, 2026.
- Plan2Map — a document-grounded geospatial agent for planning records; preprint, 2026.
Reliable AI & Safety
Capability alone does not make an AI system trustworthy. I study how to measure uncertainty, detect failures, evaluate systems when human verification is difficult, and make model and agent behaviour more stable and controllable. Much of this work starts with an existing system and adds calibration, monitoring, or practical safeguards.
- Benchmarking at the Edge of Comprehension — evaluation with humans as bounded verifiers; ICML 2026.
- Conformalized Large Language Models under Configuration Shift — calibration under changing inference settings; EMNLP 2026.
- TraceDet — hallucination detection in diffusion language models; ICLR 2026.
- BiasBusters — tool-selection bias in agents; ICLR 2026.
- Safer by Diffusion, Broken by Context — safety failure modes in diffusion language models; ICML 2026 workshop.
Statistical Foundations for AI
I develop causal and statistical approaches to understand and improve learning under shifts, interventions, and uncertainty. This includes identifying when generalisation is possible and translating structural assumptions into better learning methods.
- Causal Fine-Tuning under Latent Confounded Shift — causal identification and robust adaptation; ICML 2026.
- Structured Learning of Compositional Sequential Interventions — compositional intervention learning; NeurIPS 2024.
- Intervention Generalization: A View from Factor Graph Models — identifying outcomes under new interventions; NeurIPS 2023.