Scaling Generalization: Autonomous Intelligent Agents for the ARC-AGI Challenge

About the project

About the Project Introduction You will have complete freedom to define your research direction within the broader AI landscape: from core methods to applications, and from theory to systems. Here is a project to give you an idea of the cutting-edge directions we are actively exploring:  While Large Language Models (LLMs) and multimodal systems have demonstrated remarkable capabilities, their capacity for true abstract reasoning and broad generalization remains a critical bottleneck. Addressing these limitations, such as those highlighted by François Chollet's Abstraction and Reasoning Corpus (ARC), is essential for transitioning towards more capable, general-purpose AI systems. Concurrently, as AI systems become more autonomous, developing robust frameworks for AI safety, including the evaluation of ideological encoding and alignment, is paramount. This fully funded studentship at the University of York offers a unique opportunity to tackle these foundational challenges. The project empowers the student with complete intellectual freedom to define their own research trajectory (from core methods to applications) within the landscape of AI reasoning, generative agents, and safety. Aims and objectives The aim is to bridge the gap between deep learning and true abstract reasoning while ensuring the safe deployment of advanced AI. Specific objectives include: Investigating Reasoning: Developing novel methods for infusing reasoning capabilities and structural priors into LLMs, LVMs, and multimodal systems. Developing Agents: Designing intelligent agents capable of solving ultra-ambitious reasoning tasks, specifically targeting the ARC-AGI 2.0 and 3.0 challenges. Augmenting Performance: Exploring human problem-solving databases and collecting human performance data to augment and refine LLM reasoning trajectories. Advancing Safety: Establishing robust frameworks for AI safety, evaluating ideological encoding within frontier models, and addressing critical societal questions. Fostering Independence: Providing a supportive environment where the student can build a visionary, independent research agenda from the ground up. Methodology Due to the intellectual independence afforded by this studentship, the precise methodology will be shaped by the successful candidate. Foundational approaches are expected to include: Benchmark Evaluation: Utilising ARC as a primary diagnostic tool to evaluate model generalisation outside of their training distributions. Architectural Innovation: Fine-tuning foundation models, exploring neuro-symbolic or hybrid architectures to infuse explicit reasoning pathways. Agentic Workflows: Developing multi-agent systems and leveraging human datasets to train agents in step-by-step logical deduction. Safety Testing: Conducting rigorous empirical evaluations, including behavioural evaluations and stress-testing. Expected outcomes and impact The successful applicant will have freedom to define their research direction within the broader AI landscape: from core methods to applications, and from theory to systems. This project pushes the boundaries of broad machine generalization and robust AI safety. Primary academic outcomes will be high-impact publications at top-tier international conferences and journals, building upon the group’s established pipeline to venues such as NeurIPS and Nature Scientific Reports . Broadly, the research will impact how the AI community approaches fundamental reasoning bottlenecks and alignment. The student will emerge as an independent researcher, with funding to travel to international conferences and lead next-generation AI development. Training and support You'll receive training and guidance in research, writing and presenting skills to support your development during your PhD. You'll also cover topics such as employability skills, research management and leadership, and graduate teaching assistant training. In addition, York Graduate Research School works alongside the Department to offer high quality training, peer to peer support, professional development advice, and opportunities to engage others with your research. Location Become part of our vibrant community and contribute to inspirational and life-changing research. You will be based in the Department of Computer Science at the University of York - an exciting and welcoming hub for innovation and collaboration with a modern and inclusive working environment. In our lakeside home on Campus East, you'll benefit from world-class laboratories and collaboration spaces. The University of York is located a short distance from York city centre. Our historic city is consistently voted as one of the friendliest, safest and best places to live in the UK. Find out more about student life at York Entry requirements This funded PhD opportunity is open to individuals eligible to pay tuition fees at the UK (Home) rate. You should hold or expect to achieve the equivalent of at least a UK upper second class degree in a relevant discipline (or equivalent). We are willing to consider your application if you do not fit this profile, providing you are able to demonstrate that you have sufficient computer science knowledge and experience to succeed on the programme. We're sorry, on this occasion this opportunity is not available to international students, or to individuals who wish to study via distance learning. How to apply Please submit your application online . Please quote the project title 'Scaling Generalization: Autonomous Intelligent Agents for the ARC-AGI Challenge' in your application. Supporting documents you will need to submit with your application. More information about the application process If you have any questions about this opportunity, please email soumya.banerjee@york.ac.uk

Research areas