������

The other projects to receive support include:

Adaptation and repair in real-time conversational AI evaluation

Professor | School of Humanities, Faculty of Arts and Social Sciences; Charles Perkins Centre; Sydney Southeast Asia Centre

Professor Enfield leads a project that will enable AI agent developers to create agents that better adapt and respond to users’ language use and improve user satisfaction.

“When humans hold conversations, we are constantly analysing how the other person talks and adjusting our own language to make sure the conversation flows, we both understand each other, and we both feel comfortable,” said Professor Enfield. “Conversational AIs can analyse a human’s language use, but we don’t know how they adapt their own language use to the conversation and what impact that has on users’ satisfaction with the AI.”

The team will develop a framework to assess AI’s ‘conversational efficiency’ and ability to align with users’ speech, helping developers ensure that their AI products meet consumers’ needs and improve user satisfaction.

Quantum state certification with simple measurements

Dr | School of Computer Science, Faculty of Engineering

Dr Canonne leads a project which will support the development of scalable quantum computing systems. Quantum computers are delicate, being highly sensitive to 'noise'. The qubits they rely on (the quantum version of the 'bits' and 'bytes' regular computers use) tend to lose all useful information very quickly and turn into pure noise.

"Preparing complicated quantum 'states' – that is, configurations of many qubits we can use to perform interesting computations – is a very costly and resource-intensive process," said Dr Canonne. "So it would be nice if we could delegate this complex task to a third party, and just 'purchase' the states we need on demand. But then, how do you verify that what that third party gave you is the quantum state you asked for?"

“We aim to answer this question, certifying that a given state corresponds to its 'specifications' while using a lot fewer resources than what would be needed to create it from scratch.”

Towards characterising the reasoning-efficiency frontier in large language models

Associate Professor ����School of Computer Science,��Faculty of Engineering

Modern generative AI tools are incredibly capable, often spending extra time "thinking" through complex problems to provide better answers. However, this deep reasoning takes significantly more time, energy, and computing cost (often measured in ‘tokens’).

Associate Professor Liu leads a project which will characterise the ideal effort-to-output balance for AI and other large language models, and design a training program that teaches these models to implement this ‘reasoning-efficiency frontier’.

“Our research will help users to implement AI at scale by helping the AI technology adaptively figure out how much effort to use for each different request,” Associate Professor Liu said. “Current solutions are case-by-case specific and have to be reworked for different applications – our work will make AI more efficient across dynamic use cases.”

AI-powered proxy avatars: mediating nonverbal interaction across heterogeneous spaces for distributed mixed reality collaboration.

Professor ����School of Computer Science,��Faculty of Engineering

Mixed reality has the potential to make remote collaboration feel as natural as working side by side, by effectively 'beaming' people into each other’s spaces. But this creates a new problem: actions that make sense in one space often don’t translate well into another.

Professor Velloso’s project explores a new approach to address this challenge. Rather than forcing everyone into the same virtual environment, the research focuses on helping people feel as if remote collaborators’ actions are consistent with their own surroundings.

“We’re interested in how technology can adapt to people, not the other way around,” said Professor Velloso. “If I point to something in my space, that action should still make sense to you in yours.”

“The key innovation is in the use of AI-powered proxy avatars to interpret what people are trying to do and adjust how their actions appear to others, preserving the meaning behind gestures and movements.”