ÌìÃÀ´«Ã½

Modern AI systems increasingly operate not only as purely parametric predictors but in practice range from context augmented models to active agents: planning sequences of actions, querying external knowledge sources, and pursuing long-horizon goals across complex environments.

Our work investigates the foundations and failure modes of such agentic architectures, with a particular focus on harness and tool design, planning capabilities, model training and broader impacts of autonomous agents.

We study how agents decompose tasks, develop measures of information gain underlying tool calling decisions, and integrate structured and unstructured knowledge at inference time. A central concern is the tension between autonomy and reliability - how agents can behave robustly in open-ended settings while remaining amenable to oversight and correction.

We develop theoretical frameworks, training algorithms,  empirical benchmarks and systems to characterise agent behaviour, and examine how retrieval mechanisms affect the faithfulness, calibration and coherence of model outputs.

Keywords: agentic AI, retrieval-augmented generation, tool use, long-horizon planning, multi-step reasoning, partial observability, knowledge grounding, autonomous systems.

The development of capable AI systems rests on foundational advances in training algorithms, model architectures, and reasoning procedures. Our research addresses these foundations directly, studying how learning algorithms - from probabilistic methods to classical gradient-based supervised learning to reinforcement learning - interact with architecture design to produce systems with structured, generalisable knowledge.

We work across modalities and model families, examining diverse families of foundational models including language models and multimodal systems, graph-based architectures and hybrid neural-symbolic approaches. A core focus is developing more effective and efficient methods for modern deep learning: data-centric techniques, the geometry of loss landscapes, continual learning, and the role of scale.

On the reasoning side, we develop and analyse algorithms for inference-time computation - including test time methods such as meta-reasoning, improvements on chain-of-thought, and structured decomposition - asking when and why such procedures improve upon single-pass prediction. We also study the design of reasoning systems that go beyond pattern matching to exhibit robustness, compositionality, and systematic behaviour across novel problem distributions.

Keywords: training algorithms, optimisation, deep learning theory, model architectures, neural-symbolic systems, multimodal learning, reinforcement learning, reasoning algorithms, inference-time computation, generalisation, compositionality, foundation models.

As AI systems become more capable, ensuring that they behave in accordance with human values and remain interpretable and controllable becomes increasingly critical.

Our research addresses this challenge across several dimensions. We develop formal and empirical frameworks for evaluating model behaviour - including robustness to adversarial inputs, consistency across contexts and alignment with specified objectives. We also investigate techniques for eliciting and representing human preferences correctly (including conflicts under competing demands) and examine how uncertainty in value specifications propagates through training and deployment.

Given the proprietary cost of AI systems, we develop method allowing effective value re-alignment of systems to individual actors without the need for large-scale resources, democratising access to the ethics of foundation models. A key theme is the development of evaluations that are both rigorous and practically informative, enabling principled comparisons across models and deployment settings.

Keywords: AI safety, alignment,  robustness, adversarial evaluation, value learning, preference elicitation, model evaluation, interpretability, controllability scalable oversight.

The deployment of large-scale AI systems entails consequences that extend well beyond individual interactions and the topics typically studied in AI research labs - reshaping labour markets, information ecosystems, epistemic norms and distributions of power. Our research examines these second-order effects with rigour and breadth, drawing on methods from law, computational social science, economics and political science. 

We study how AI systems affect the production and spread of information, including misinformation and persuasion at scale, as well as questions of fairness, representation, and differential impact across populations. We investigate the use and limits of AI in high-stakes decision making, including the use of law, both in terms of professional use and fundamental model limitations leading to failure cases like hallucinations.

We are particularly interested in the medium-long term consequences of modern AI on downstream societal outcomes, including questions of access to justice, social justice, job displacement, productivity and economic growth, and political polarisation. This work aims to produce both descriptive accounts of real-world AI impacts and normative frameworks for evaluating them.

Keywords: AI and society, Legal AI, misinformation & censorship, AI labour impacts, distributional consequences, computational social science, model alignment and values.