Speakers

Morning session (11:00am–1:00pm)

1. Introduction: Joan Lasenby (11:00am–11:10am)

2. Keynote: Kate Knill (11:10am–11:40am)

What can we learn about a learner from speech technology … and what can’t we (yet)?

Three key aspects of a CALL system are i) the measurement of a learner's proficiency level, ii) determining which aspects of the language they do well and which need improvement and iii) providing advice to a learner on how to correct their mistakes. This talk addresses these 3 aspects applied to speaking by second language (L2) English learners. Automated holistic assessment of monologic speech, both read and free ('open’) style, is well established with several L2 English tests utilising automated scoring. The talk will describe the typical auto-marking systems that can provide an assessment for placement and/or to judge progress. For learning purposes, analytic assessment is more informative, that is assessment of individual aspects of speech, such as fluency, pronunciation and grammatical complexity. The advent of LLMs has allowed these aspects to be automatically evaluated at an even finer-grain such as what vocabulary a speaker is using and the levels at which they are doing so. Measurements of e.g. grammar usage and correctness and pronunciation choices can provide useful information to a CALL system to build a learning strategy for the learner. The approaches adopted will be described in the talk. It will also look at where there are still partially or unanswered questions such as in dialogic auto-marking and pronunciation error detection.

3. Erez Li (11:40am – 12:00pm)

Steering phage infection outcome with physiology-aware combination therapy design

Single-cell heterogeneity is often viewed as a source of noise in antimicrobial response. Here, we show that variation in phage infection outcomes is instead organised by measurable host physiological state variables. Combining time-resolved single-cell infection assays with mechanistic modelling, we find that bacterial cell-cycle position structures the distribution of ribosome abundance and cell surface area, while ribosome concentration remains tightly regulated. These variables define a predictive physiological state space for phage infection: surface area controls adsorption probability, whereas total ribosome abundance constrains viral production. We then use this framework to control infection dynamics. β-lactam treatment perturbs the host state distribution towards larger, ribosome-rich cells, simultaneously increasing adsorption and phage production. This state shift predicts, and experimentally explains, accelerated population eradication by phage–antibiotic combinations. Although similar combinations can be identified by high-throughput screening, our approach derives them from a quantitative model linking single-cell physiology to population-level outcome. These findings establish host physiological state as a mechanistic determinant of phage infection and illustrate how structured heterogeneity in living matter can be predicted, perturbed, and exploited for control.

4. Lu Chen (12:00pm–12:20pm)

World Models for Humanoid Spatial-Behavioral Intelligence

Humanoid intelligence requires more than recognizing the present: an agent must anticipate how the world will evolve under its actions and use those predictions to guide future behavior. This talk examines world-state representation as a central challenge in building such predictive agents. Starting from egocentric perception and 3D human motion, I will discuss how joint-embedding predictive learning across visual and motor modalities can unify visual understanding, action generation, and future-state prediction within a shared semantic space. Grounding these representations in metric 3D geometry makes the learned dynamics robust to changing viewpoints and provides meaningful objectives for action planning. Moving beyond world representations based on semantics and geometry, in which motion is modeled only implicitly through next-state prediction, 3D point trajectories represent motion explicitly by establishing correspondence across time, thereby aligning more naturally with physical dynamics. Together, these directions outline a path from joint visuomotor prediction to spatiotemporal dynamics modeling, toward humanoid agents that can understand, anticipate, and purposefully interact with the world.

5. Youjing Yu (12:20pm–12:40pm)

Learning without forgetting: How plasticity and dynamics enable continual learning

Animals can continuously acquire new skills and retain the ability to deploy them when needed throughout their lives. Artificial intelligence (AI) agents, in contrast, often struggle to retain previously acquired skills when learning in new contexts—a phenomenon known as catastrophic forgetting (CF). Understanding and mitigating CF remains a central challenge in both machine learning and neuroscience. Previous work has largely focused on two complementary mechanisms for mitigating CF at the network level: regulating synaptic plasticity or relying on network dynamics. Here, we hypothesize that it is the structured interaction between plasticity and dynamics that enables continual learning. To test this idea, we develop a multi-area network in which a small number of functionally distinct modules divide the labour of learning, representing different brain regions and their associated functions while a specific set of weights undergo plastic adaptation. By meta-training this network with SOFO, a memory-efficient second-order optimizer developed in lab, we obtain an optimal continual learner within this model class. The resulting network mitigates catastrophic forgetting while reproducing behavioural phenomena observed in humans, including anterograde interference and savings. Analysis of the individual modules further reveals the emergence of abstract representations, as well as distinct mechanisms supporting the retention and rapid reacquisition of previously learned skills. Together, these results suggest that metaplasticity can provide a computationally simple and biologically grounded mechanism for continual learning, while offering insights into how neural systems may preserve and reuse knowledge across changing contexts and generating testable hypothesis for experiments.

Lunch (12:40–2:00pm)

Afternoon talks (2:00pm–4:00pm)

1. Keynote: Maurice Filo (2:00pm–2:30pm)

From Genetic Feedback Control to AI-Designed Biomolecular Networks

How can we turn a desired biological behaviour into a network of chemical reactions? I will begin with our work on biomolecular feedback controllers, showing how principles from feedback control theory can be translated into genetic circuits that regulate cellular outputs despite disturbances. These implementations motivate a broader challenge: finding useful reaction networks across a vast space of possible mechanisms. The main part of the talk introduces GenAI-Net, a generative AI framework for automated biomolecular network design using reinforcement learning. Given a desired behaviour, a learning agent proposes reaction network structures and kinetic parameters, while simulations evaluate each candidate. I will show how this design loop yields diverse solutions for dose responses, logic functions, oscillations, and robust perfect adaptation, including stochastic controllers that reduce output noise. The resulting reaction networks reveal recurring design motifs and provide candidates for future genetic implementation.

2. Liam Hallinan (2:30pm–2:50pm)

Decentralised Plug-and-Play Stability Conditions for AC Grids

Power systems are undergoing a profound transformation. As synchronous generators are replaced by inverter-based resources, the grid is becoming highly distributed and heterogeneous. Classical small-signal techniques, such as eigenvalue analysis of a fixed system-wide model, are no longer scalable. Furthermore, device models are often only available as black boxes, and the network itself changes as new resources connect. Motivated by these challenges, I will present a decentralised framework that certifies the stability of the entire grid from the bottom up. The framework consists of local specifications on the frequency response of grid subsystems, which together form a grid code or interconnection protocol. If every device in the network satisfies them, these constraints collectively ensure that the multivariable Nyquist criterion holds and that the grid is stable, enabling plug-and-play operation as new devices connect. Constraints can be applied at the bus level or to extended subsystems that couple a bus to its connecting lines and loads, and different conditions can be mixed across frequency ranges, reducing conservatism compared with existing methods. 

3. Henry London (2:50pm–3:10pm)

Investigating non-monotonic feedback of host physiology on bacteriophage infection

Rapidly growing bacterial cells are generally assumed to be ideal phage hosts because they provide abundant biosynthetic resources. We find that the host physiological state creates two opposing barriers to infection success. Some cells become invisible because they lack receptors, whereas others become unproductive because excessive infection perturbs physiology. Consequently, phage success emerges from a balance between encounter and exploitation rather than either process alone.

4. Ares Panos (3:10pm–3:30pm)

Axolotl: Generation of Area Forecast Discussions From AI Weather Models

AI weather models increasingly produce high-dimensional forecasts spanning many variables, atmospheric levels, locations and lead times. For downstream use, these numerical forecasts often need to be interpreted and transformed into concise, application-relevant information such as area forecast discussions (AFDs). Large language models (LLMs) provide a natural mechanism for this transformation, but require an effective means to connect to the rich forecast state produced by the weather model. We introduce Axolotl, a parameter-efficient framework that proposes to directly couple the internal latent representation of a pretrained AI weather model to an LLM to generate AFDs. Both model backbones remain frozen, with only lightweight projection and adaptation parameters learned. We compare this latent interface with alternatives based on text summaries of selected variables and 2D gridded fields output by weather models. We train and evaluate our system using more than one million paired weather states and forecast targets derived from National Weather Service Area Forecast Discussions. The latent coupling achieves the strongest aggregate performance on a multiple-choice meteorological interpretation task, reaching 89.3% accuracy on a held-out year and 83.2% on unseen regions. It also performs strongly on open-ended forecast generation where nearly 60% of the generated forecasts receive a score of 6 or 7 from an LLM-as-a-judge, while a learned weather-state–text evaluator classifies 94.5% of the generations as consistent with the provided weather state.

5. Term Welcome Address: Joan Lasenby (3:30pm–4:00pm)