The main goal of the conference is to foster discussion around the latest advances in Artificial Intelligence

Agentic AI for Regional Analysis of Aerosol Absorption and Surface Temperature
Abira Sengupta, University of Otago

Understanding the geographical relationship between Absorbing Aerosol Optical Depth (AAOD) Li et al. [2021] and Surface Air Temperature (K) is important for investigating aerosol-climate interactions. However, analyzing large multidimensional environmental datasets remains computationally challenging. An autonomous multi-agent AI system, Agentic AI for Regional Analysis (Figure 1), is presented to coordinate end-to-end climate analytics using NASA Giovanni-derived AAOD Data fusion, quality control, spatiotemporal statistical modelling, association discovery, forecasting, and explainability are all carried out by specialised agents. The Quality Agent reduced the percentage of missing AAOD cells from 20.14% to 0.34% by using temporal linear interpolation and automatic missing-data recognition. Different spatial patterns of aerosol absorption and temperature were revealed by the Spatial Agent, which found 147 AAOD hotspots concentrated in the northern mountainous region (Figure 2(a)) and 132 temperature hotspots concentrated primarily in the southern lowlands (Figure 2(b)). AAOD and temperature relationships vary across spatial and temporal dimensions, as evidenced by the Relationship Agent’s identification of localised lag interactions, positive temporal co-movement (r = 0.589, p < 0.001), and a substantial negative between-location spatial relationship (r = −0.809, p < 0.001). With 1,298 locations (88.2%) displaying warming paths and 916 locations (62.3%) displaying declining AAOD paths (Figure 3), the Forecasting Agent fitted site-specific Holt-Winters Exponential Smoothing models (S = 12) Winters [1960], indicating spatially heterogeneous forecast patterns. While the Report Agent combined the multi-agent outputs into an organised analytical summary, the Scientific Critic Agent offered an automated protection against causal overinterpretation. All things considered, the architecture shows how Agentic AI can be used for autonomous, scalable, and repeatable spatiotemporal environmental investigation. 

From Synthetic EEG to Real Source Localisation: Combining Physiologically Informed Data Generation and Leadfield-Guided Learning 
Ade Nurhopipah, University of Otago 

Implementing deep-learning models for EEG source localisation requires large amounts of labelled training data, yet ground-truth source locations are rarely available for real EEG recordings. We present an integrated synthetic-to-real framework combining physiologically informed EEG generation with leadfield-guided deep learning. Five subject-specific finite-element head models and their corresponding leadfields were used to generate synthetic datasets by projecting parameterised event-related source activity with random dipole orientations onto scalp sensors. To approximate the spatial structure of real EEG, the simulated responses were mixed with correlated background activity generated from each subject’s pre-stimulation EEG covariance matrix across a range of signal-to-noise ratios. Each synthetic dataset was used to independently train an ESoLo model—an attention-based network supervised by Gaussian spatial labels and a leadfield-topography loss that preserves the physical source-to-sensor relationship during learning. The pipeline was evaluated on real cortico-cortical evoked potential recordings, using intracranial stimulation coordinates as spatial references. It achieved mean localisation errors of 23.70 ± 12.38 mm under matched head anatomy and 34.32 ± 17.13 mm in zero-shot evaluation across 31 subjects with unseen head anatomies. Leadfield-informed supervision reduced the cross-subject error by 6.96%. These findings demonstrate how subject-specific synthetic data generation and physics-guided learning can be combined to support EEG source localisation when labelled real data are limited. 

Seeing Less, Understanding More: Human-Vision-Guided AI Optimisation for Embedded Machine Vision 
Anastasia Mozhaeva, Eastern Institute of Technology (presented by Amr van den Adel)

Modern machine vision systems process enormous amounts of visual data, although much of it contributes little to AI decision-making. Human vision works differently, processing the visual field with varying spatial and temporal detail. This research asks a simple question: can artificial intelligence become more efficient by learning what not to see? We introduce a perception-aware AI optimisation approach based on experimentally measured characteristics of the human visual system. By reducing perceptually redundant information before inference, computer vision models can focus computation on the data most relevant to perception and action.

Experiments on embedded AI platforms, including NVIDIA Jetson, show that this approach can substantially reduce the amount of visual data processed while maintaining, and in some cases improving, the performance of existing computer vision models, including YOLO-based object detection. Current work extends the concept to resource-constrained microcontrollers, where memory, computation and energy are critical limitations. Rather than building larger models to process more data, we explore AI systems that selectively process the right data, with applications in autonomous robots, drones and other edge-AI systems. 

Explainable, interdisciplinary AI for Aotearoa 
Andrew Lensen, Victoria University of Wellington 

My Applied and Explainable AI (AXAI) group builds AI systems that people can understand and apply them to problems that matter to Aotearoa New Zealand. We work widely across disciplines with ecologists, legal and policy scholars, cognitive scientists and software engineers.

In this talk I will give an overview of our work. Examples will include recognising individual kākā for conservation management, modelling forest change and flooding, testing whether machine learning can explain if two pieces of Java bytecode behave the same, and analysing whether NZ’s laws can effectively control deepfake harm.

I will finish the talk by discussing the problems that these projects share, including how we evaluate an explanation with the people who will use it, and how we keep data and models under local control here in Aotearoa. 

The Representation of Socio-Cognitive Trust in LLMs 
Anushka Debnath, University of Otago 

Trust plays a central role in human collaboration and decision-making, yet little is understood about how large language models (LLMs) internally represent trust-related concepts. In this work, we investigate trust in LLMs through behavioural and latent-space analysis grounded in the socio-cognitive trust theory of Castelfranchi and Falcone. We evaluate Gemma-3-27b-it, a state-of-the-art open-weight LLM, across multi-stage interpersonal scenarios involving trust and its five socio-cognitive factors. Using the resulting model assessments, we fit linear regression models to examine the relationship between trust and its factors. We further examine whether similar structures exist within the latent space of Gemma-3-27b-it using existing concept extraction methodologies. By combining behavioural regression coefficients with extracted concept vectors, we construct predicted trust representations and compare them with directly extracted trust vectors using cosine similarity to see whether trust is encoded in a structured and linearly recoverable form within LLM latent spaces. 

Toward Objective Monitoring of Spinal Cord Injury using Electrophysiological Signals 
Aron Jeremiah, University of Auckland 

Spinal cord injury (SCI) lacks objective and scalable methods for tracking neurophysiological recovery, with current assessments relying on specialist interpretation, imaging with limited longitudinal sensitivity, or functional scoring that does not capture underlying neural change. This limits both clinical decision-making and the evaluation of emerging therapies.

We develop an interpretable representation-learning pipeline that constructs structured latent spaces from spinal cord electrophysiological signals. By combining mathematically grounded times-series feature transformations with ensemble modelling, the approach enables direct inspection of the features and visualisation of the feature-space geometry used to distinguish between neural recovery states.

Applied to invasive recordings from healthy and contused spinal cords, the method reveals separable low-dimensional structure (AUC = 0.8) using a small subset of time-series features. These representations enable downstream interpretation and hypothesis generation about underlying physiology and potential recovery mechanisms

This work demonstrates that spinal cord recovery state can be objectively characterised from neural data, providing a pathway toward interpretable monitoring of injury and recovery. This capability has the potential to support evidence-based continuation of rehabilitation and accelerate the development and evaluation of therapies targeting neural function. 

Intervention Complexity as a Canonical Reward and a Measure of Intelligence 
Brendan McCane, University of Otago 

The Legg--Hutter universal intelligence measure assesses general intelligence as expected reward across all computable environments, weighted by simplicity. The measure presupposes an externally specified reward function, raising the question of whether a canonical choice exists. I propose intervention complexity as a canonical reward, prove that it is canonical subject to mild conditions, and subsequently propose a two-dimensional characterisation of intelligence based on agent competence and learning efficiency. 

Tabular Foundation Models for Wildfire Risk Prediction 
Brendon Woodford, University of Otago 

Wildfires increasingly pose a major threat to the environment as evidenced by the recent extreme wildfire events in the Northern Hemisphere. Artificial Intelligence (AI) and Machine Learning (ML) have emerged over the past decade to support decision-making around how to forecast and manage these natural disasters. For example, predicting the onset of wildfires or the amount of damage to the landscape a wildfire might cause.

Training wildfire prediction models involve tabular data that records longitudinal data acquired on weather, fuel, vegetation types and topography for a specific region. However, when there is a lack of this longitudinal data, generating a ML model to predict wildfire risk effectively becomes very challenging. Recently there has been work in applying Tabular Foundation Models (TFM) for the purpose of leveraging in-context learning for predicting outcomes from tabular data as TFM are modelling casual information and are especially robust to outliers and missing values.

We report research that compares TFM models for predicting the risk of wildfire in different regions of Aotearoa New Zealand with ML models such as XGBoost. Contributions from this study demonstrate the utility of TFM as a viable model for predicting the risk of wildfire. 

Sustainable and explainable AI through decomposed-error-driven AutoML 
Caitlin Owen, University of Otago 

Machine learning (ML) is booming in many domains, attracting non-expert users. However, selecting an algorithm is time-consuming and difficult, even for experts. This boosts the appeal of AutoML, a field of AI that involves the data-driven evaluation and selection of a pipeline of algorithm modules, comprising a module for each algorithm design choice (e.g., number of hidden layers for neural networks and crossover method for genetic programming). AutoML is making predictive models more accessible, driving their use for critical decisions. While it can deliver good predictive performance, AutoML is resource and energy-intensive, due to the evaluation of many ML pipelines. The computing power required for AI, particularly deep neural networks (DNNs), is doubling approximately every six months. AI data centres also consume vast amounts of water. Thus, AutoML and AI priorities must include sustainability. Currently, AutoML minimises aggregate (total) prediction error, which lacks transparency on module selection and conflicts with explainable AI, crucial in high-risk domains (e.g., diagnosis). This adds another layer of ‘black-boxness’ to ML. Despite a shift to human-centred AutoML, total-error-driven AutoML masks model behaviour and limits customisation. As a recipient of the Royal Society Te Apārangi Mana Tūāpapa Future Leader Fellowship, I will develop novel AutoML methods guided by splitting total error into underlying error components. This targeted error reduction is expected to provide greater explainability and require less energy through more efficient ML algorithm search. It will also enable the prioritisation of error components to provide domain customisation and greater trust. As the standard bias and variance decomposition does not distinguish the different sources of variance for stochastic algorithms, I will use an error decomposition framework that I developed that distinguishes between variance due to the algorithm and variance due to the training data. 

Beyond Keywords: Fine-Tuned Language Models for Māori-Relevant Research Metadata Classification 
Dinushi Thathsarani Jayasinghe, University of Waikato 

Institutional repositories in Aotearoa New Zealand contain research records with substantive Māori relevance, but this relevance is not always represented in a dedicated metadata field. Traditional lexical methods such as TF-IDF rely on patterns of word occurrence, which can be limiting because Māori-related terms may appear in records where Māori are not a substantive focus. Relevant records may also be identified through combinations of information across titles, abstracts, and subject fields rather than through a single keyword. This study investigates whether contextual language models can improve the identification of Māori-relevant research metadata.

We used 16,589 publicly available metadata records from the University of Waikato Research Commons. Māori relevance was defined at the record level using an operational framework covering Māori people and communities, culture and language, knowledge systems, history and governance, land and environmental resources, material culture, and Māori-specific research or data contexts. Incidental Māori terminology, place names, passing references, and minor Māori subgroups within broader studies were treated as non-Māori-relevant. Computational reference labels were generated predominantly using Gemini 2.5 Flash. We compared a TF-IDF logistic regression baseline with EmbeddingGemma-300M before and after triplet fine-tuning and Gemma-3-4B using zero-shot and LoRA fine-tuned classification.

Evaluation on a held-out test set of 3,318 records showed that task-specific fine-tuning improved performance. The TF-IDF baseline achieved a macro F1 of 0.8721. Fine-tuned EmbeddingGemma-300M reached 0.9265, while LoRA fine-tuned Gemma-3-4B achieved the highest macro F1 of 0.9343. The fine-tuned EmbeddingGemma model also achieved higher recall than the fine-tuned Gemma-3-4B model, showing that the smaller model remained competitive for repository screening.

The results indicate that word-level matching alone is insufficient for identifying Māori-relevant research metadata. The fine-tuned models improved the identification of Māori-relevant records and can be used to flag records for further review. The system is not intended to make authoritative judgements about Māori identity, provenance, cultural authority, or governance. Because the reference benchmark was predominantly model-generated, the findings require validation with Māori and domain experts. 

Tackling Modern Operational Challenges: Batch Shops, Green Scheduling and Lifelong Learning 
Fangfang Zhang, Victoria University of Wellington 

This talk provides a rapid overview of four distinct, cutting-edge research tracks in operational optimization and artificial intelligence. First, we explore recent advancements in batch-based job shop scheduling, focusing on strategies to manage complex grouped operations and maximize throughput. Second, we examine energy-saving scheduling through a multi-objective technique, discussing frameworks that successfully balance strict production targets with the critical need to minimize energy consumption. Third, we introduce lifelong learning which is an emerging machine learning technique designed to help models continuously acquire new knowledge over time without forgetting past tasks. Finally, we step out of the theoretical space and into a real-world maritime application, highlighting our practical work in tugboat scheduling to optimize vessel logistics in dynamic harbour environments. Together, these highlights showcase the diverse approaches required to solve today's most demanding scheduling and operational challenges. 

Combining GNSS Reflectometry and Radar for Earth Observation 
Felix Marattukalam, University of Auckland 

No single sensor can capture every environmental change at the detail and frequency needed for timely decisions. GNSS Reflectometry (GNSS-R) uses reflected navigation signals to detect changes in land and water, while Synthetic Aperture Radar (SAR) produces detailed images across wider areas, even through cloud and at night. Together, they offer a richer view than either can provide alone.

This talk explores how these complementary measurements can be combined to monitor soil moisture, flooding and coastal change. Using examples from New Zealand’s Rongowai instrument and satellite missions, we will discuss how data fusion and physical modelling can help fill gaps between observations and improve the reliability of environmental information. The broader goal is to turn diverse Earth observation data into more timely and useful insights for monitoring and responding to change. 

Resource-Efficient AI: Real-Time Anomaly Detection at the Edge 
Guilherme Weigert Cassales, Massey University 

Aotearoa New Zealand’s strategic AI priorities demand lightweight, energy-efficient intelligence tailored to primary production, environmental monitoring, and resilient infrastructure. Traditional cloud-based machine learning often fails in remote settings due to high latency, bandwidth limitations, and compute costs. This talk presents adaptive streaming isolation models as a resource-efficient paradigm for real-time anomaly detection. By processing streaming sensor data directly at the edge, these lightweight algorithms handle dynamic environmental shifts without heavy compute requirements. We discuss energy-efficient stream learning methods and show how open-source frameworks like CapyMOA empower local researchers and industries to deploy sustainable, real-time AI. 

Signer-Independent Medical Sign Recognition from Skeletal Motion Using BiGRU with Temporal Attention 
Hansi Gunasinghe, Southern Institute of Technology 

Communication barriers between healthcare professionals and patients who are deaf or hard of hearing can affect the effective exchange of health information. This study investigates signer-independent recognition of medical signs from skeletal motion sequences using the MedSLSet multimodal medical sign dataset. Three-dimensional skeletal coordinates were represented as fixed-length temporal sequences and evaluated using recurrent neural network architectures based on bidirectional gated recurrent units (BiGRUs). Particular attention was given to the effect of temporal attention and temporal augmentation on recognition performance. Under a signer-independent evaluation setting, the baseline BiGRU achieved an accuracy of 67.6% and a macro-F1 score of 65.4%. Incorporating temporal attention substantially improved performance, achieving a mean accuracy of 78.3% ± 2.9% and macro-F1 of 77.2% ± 3.3% across five runs, while temporal augmentation reduced performance. The findings indicate that explicitly learning the relative importance of temporal regions can improve skeletal medical-sign recognition while preserving the temporal structure of sign movements. A proof-of-concept downstream pipeline further maps recognised medical signs to structured healthcare-oriented summaries, illustrating a potential pathway from sign recognition toward AI-assisted healthcare communication. 

Towards Developing a Culturally-Appropriate Speech Emotion Recognition System for te reo Māori 
Himashi Rathnayake, University of Auckland 

Emotions play an important role in human communication, providing insight into an individual's mood, goals, and motivations. With the increasing integration of speech-based AI technologies into everyday communication, there is growing interest in Speech Emotion Recognition (SER), the automatic identification of emotions from speech. However, SER research has primarily focused on well-resourced languages, leaving languages such as te reo Māori, the only official Indigenous language of New Zealand, underrepresented. A further challenge in SER is the reliance on a limited set of predefined emotion categories, often based on six basic emotions or their direct English translations. The universality of these categories remains debated due to cross-cultural differences in emotion conceptualisation, as these theories were not developed using culturally grounded approaches. Māori emotion studies have shown that Māori emotional expressions can differ from those of New Zealand Europeans, and that important nuances can be lost when translated into English.

This research investigates a culturally grounded approach to SER for te reo Māori. We first explored emotions expressed in Māori community media through a questionnaire and focus group discussions with Māori participants, identifying 16 emotion categories relevant to te reo Māori. Based on these findings, we developed Pā-Kākare, the first known emotional speech corpus for te reo Māori, comprising 3,840 utterances recorded by professional Māori-speaking voice actors. Acoustic analysis showed that features including fundamental frequency, intensity, speech rate, and glottal-source characteristics significantly differentiated most emotion categories. We subsequently evaluated these features using machine learning approaches to develop a Māori SER system. The current system identifies 16 emotion categories with approximately 60% accuracy, demonstrating the feasibility of developing culturally informed speech emotion recognition systems for Indigenous and low-resource languages. 

Beyond More Data: Linguistically Informed Speech AI
Jesin James, University of Auckland 

More data is not always the answer. For many languages, it is not even an option. Speech technology development is increasingly data-driven, placing languages with limited training data, such as New Zealand English and te reo Māori, at a disadvantage. Our research explores how linguistic knowledge can be integrated into the training and evaluation of speech models, thereby moving beyond conventional data-driven approaches and aggregate performance metrics. We specifically investigate how linguistic analysis of vowel sounds in accents of English can identify specific speech sounds and patterns that are poorly modelled. We then examine how these modelling errors relate to human perception of the model outputs.

We will present our work so far towards developing linguistically informed approaches to text to speech synthesis model evaluation and training, with the longer-term goal of identifying and addressing bias during model training rather than after deployment. This provides a pathway towards speech technologies that make better use of existing linguistic knowledge and perform more equitably across languages and varieties.

Active Learning Through a Data-Centric Lens 
Katharina Dost, University of Canterbury 

Data-centric machine learning shifts attention from improving models to improving the data they learn from. In this talk, I will explore what this perspective might mean for active learning.

Active learning is typically used to select additional observations that improve predictive performance as efficiently as possible. But in many applications, performance alone is not enough: we also want models that are reliable under imperfect coverage, sampling bias, distribution shift, or other weaknesses in the available data.

I will discuss how active learning could be broadened to incorporate such reliability considerations, and how a more data-centric view of data acquisition may help us build models that are not only more accurate, but also more robust. 

Approximating Many-Body Quantum States Using Deep Learning 
Krystof Krsek, University of Otago 

Solving the quantum many-body problem is fundamentally limited by the exponential growth of the Hilbert space with system size. An exact treatment requires knowledge of the many-body wave function itself, and even efficient stochastic approaches such as quantum Monte Carlo (QMC) can fail as system size and correlation strength grow. Neural networks offer a compact and expressive way to approximate such wave functions directly, and the resulting approximation can additionally serve as a guiding wave function for importance sampling, improving QMC performance in these harder regimes.

We implement neural network ansätze for the many-body wave function in Julia and train them with variational Monte Carlo (VMC), a physically motivated, self-reinforcing scheme that minimizes the variational energy. We use the open-source package Rimu.jl [1] as the interface to the physical systems, and employ projector Monte Carlo within the same framework to validate the trained wave functions against independently obtained results.

The goal of this work is to build a general framework for testing different neural network architectures as many-body wave function ansätze and to systematically assess their performance on quantum systems. So far, we have implemented fully connected neural networks (FCNNs) and are extending the framework to convolutional neural networks (CNNs), with graph neural networks (GNNs) and transformer-based architectures planned. A central question throughout is how the physical symmetries of the system can be built directly into the network architecture rather than left for the training to learn.

Ongoing work includes comparing architectures in terms of accuracy, training stability, and transferability across system sizes, and extending the framework toward larger and more strongly correlated systems, where conventional QMC methods face the greatest difficulties.

[1] M. Čufar, C. J. Bradly, R. Yang, E. Pahl, and J. Brand, “Rimu.jl: Random integrators for many-body quantum systems,” (2026), arXiv: arXiv:2601.19505. doi: 10.48550/arXiv.2601.19505.

All and Only the Referents: Counterfactual Evidence Boundary Learning for Generalized Visual Grounding 
Lysa xiao, University of otago 

Generalized visual grounding must localize all and only the objects named by an expression, including when no target exists. In object-centric models, this requires an exact-set decision over object hypotheses. Yet prevailing methods learn that decision through candidate-wise scores, target counting, or existence prediction—partial signals that do not determine whether the selected hypotheses collectively form the correct set. A prediction can have the right cardinality and individually plausible members while still omitting a referent or including a distractor. We introduce Counterfactual Evidence Boundary Learning (CEBL), which aligns supervision with the final decision. CEBL scores the complete selected excluded membership state and learns counterfactual evidence boundaries by ranking valid referent sets above hard counterfactuals encoding omission, addition, substitution, and existence errors. At inference, typed fixed-budget search finds a high-scoring state without enumerating all subsets. The same membership decision serves both Generalized Referring Expression Comprehension (GREC) and Generalized Referring Expression Segmentation (GRES). For GRES, Box-Conditioned Shape Realization (BCSR) removes redundant position and scale prediction from shape reconstruction: predicted boxes provide the spatial frame, compact DCT codes represent only box-local shape, and CEBL selects the contributions to merge. On the gRefCOCO validation split, our unified model achieves 81.25 Pr@F1 for GREC and 82.57 gIoU for GRES, outperforming the strongest reported baselines by 7.75 and 9.21points, respectively. We will open-source our code. 

Reinforcement Learning-based Home Energy Management with Heterogeneous Batteries and Stochastic EV Behaviour 
Meng Yuan, Victoria University of Wellington 

The widespread adoption of photovoltaic (PV), electric vehicles (EVs), and stationary energy storage systems (ESS) in households increases system complexity while simultaneously offering new opportunities for energy regulation. However, effectively coordinating these resources under uncertainties remains challenging. This paper proposes a novel home energy management framework based on deep reinforcement learning (DRL) that can jointly minimise energy expenditure and battery degradation while guaranteeing occupant comfort and EV charging requirements. Distinct from existing studies, we explicitly account for the heterogeneous degradation characteristics of stationary and EV batteries in the optimisation, alongside stochastic user behaviour regarding arrival time, departure time, and driving distance. The energy scheduling problem is formulated as a constrained Markov decision process (CMDP) and solved using a Lagrangian soft actor-critic (SAC) algorithm. This approach enables the agent to learn optimal control policies that enforce physical constraints, including indoor temperature bounds and target EV state of charge upon departure, despite stochastic uncertainties. Numerical simulations over a one-year horizon demonstrate that the proposed framework reduces the total annual operating cost by 7.9% to 17.0% relative to four learning-based comparison methods and by 33.0% relative to a perfect-foresight model predictive control baseline, while satisfying the EV departure requirement on every day of the test year.

This work is joint with Ye Wang, Xinghuo Yu, Torsten Wik and Changfu Zou, and has been published in IEEE Transactions on Sustainable Energy (DOI: 10.1109/TSTE.2026.3719940). 

Unsupervised Detection of Volcanic Unrest using Transformer based Rectified Flow 
Nancy Chugh, Victoria University of Wellington 

Volcanic eruptions yield many hazards that can risk life, infrastructure, and other essential services. Most eruptions are preceded by volcanic unrest, which is expressed by changes in seismic behavior, ground deformation, gas emissions, and other monitoring signals. Timely detection of volcanic unrest can enable enhanced hazard assessment and decision making. However, the signals and precursors associated with unrest have complex, non-linear behavior, which is variable across eruptions and different volcanoes, they also rely upon labels generated by volcanologists that are susceptible to bias and non-uniqueness.

Recent studies demonstrated that machine learning has a potential for automated volcanic activity detection and short-term eruption forecasting. However, current studies generally focus on specific forms of volcanic activity or particular signal types, and many rely on labelled data. So, there is still a gap of a model that use multi-modal data streams like seismic, ground deformation and SO2 flux to detect volcanic unrest using unsupervised approach.

This study proposes an unsupervised, multivariate reconstruction-based anomaly-detection framework that will use multiple types of data streams, not just seismic. The primary objective is to address the limitations of traditional time-series anomaly detection models for detecting volcanic unrest—which either rely on supervised learning or rely on rigid pointwise reconstruction errors (L2/MSE). The key idea is that the generative model (RFFormer) serves as a normal reference generator that reconstruct how the signal would look if the system were operating normally, anomaly scores are then derived by calculating the Wasserstein divergence. The model employs a Conditional-Aware Encoder and a Temporal Decomposition Decoder that will help Rectified Flow to model clean time-series dynamics. Instead of directly minimizing only the pointwise L2 errors between observation and reconstruction, the model will minimize the Wasserstein distance. This formulation effectively captures structural deviations, baseline drifts, temporal dynamics and complex multi scale patterns, which couldn’t be possible with pointwise L2 error.

Comparative Analysis of LLM Interpretation: Assessing Cultural Contextualisation in Māori Leadership Concepts. 
Natalia Ortega, ICL Graduate Business School 

The rapid deployment of large language models (LLMs) presents new challenges for the accurate representation and contextualisation of complex, culturally grounded knowledge systems. The study addresses a critical gap by examining how four freely accessible AI platforms, ChatGPT, Gemini, Claude, and Grok, interpret five core Māori leadership concepts drawn from established scholarship. Our central objective is to assess the fidelity, depth, and cultural grounding inherent in AI-generated interpretations of Te Ao Māori (the Māori worldview).

Methodologically, we employ a comparative qualitative analysis framework built upon an analytical rubric developed specifically for assessing Indigenous knowledge. This rubric moves beyond simple factual accuracy to evaluate deeper criteria: conceptual contextualisation, the presence or absence of cultural reductionism, and the identification of unsupported cultural claims. By systematically prompting multiple LLMs with the same foundational Māori concepts, we generate a comparative dataset that allows us to map patterns of interpretation, bias, and knowledge appropriation.

We anticipate that our findings will reveal distinct challenges modes in generative AI when handling highly contextualized Indigenous epistemology. The study aims to demonstrate how reliance on generalist AI platforms risks producing interpretations of cultural knowledge. The research holds significance for both the fields of AI ethics and Indigenous scholarship in Aotearoa New Zealand. Furthermore, this work contributes a vital Māori perspective to global discourse on generative AI, advocating for methodologies that support cultural partnership in technological advancement.

Towards Open-Set Re-Identification of Kākā 
Nicolas Samelson, Te Herenga Waka—Victoria University of Wellington 

The kākā (Nestor meridionalis) is a taonga species of Aotearoa New Zealand. In 2002, this parrot was reintroduced to Zealandia, a fenced sanctuary in Wellington. Its population has since grown and spread beyond the fence, exposing birds to new threats such as brodifacoum poisoning from urban pest control. Individual-based monitoring is vital to understand the impact of these threats, yet only a small fraction of birds is individually marked in Wellington. Automated re-identification (Re-ID) enables recognition of kākā by appearance alone, across varying capture conditions. However, animal Re-ID typically evaluates models on open-set datasets at a fixed ratio of unseen individuals; conditions that do not occur in growing wild populations. Additionally, models are rarely tested within end-to-end systems, where detection and tracking errors propagate to identity matching. This research presents a pipeline to curate a time-, location-, and encounter-aware open-set dataset for kākā Re-ID. Out of 28 recording days spanning two years, the first day (222 videos) is manually annotated with boxes, tracks, species, and band identities. Detection, species classification, and tracking models are evaluated per stage and in cascade to assess whether semi-automation can scale to the full dataset and feed an automated Re-ID system. Splits are then built across days, locations, and identities to vary the openness ratio, the proportion of unseen individuals, simulating a growing population. Preliminary results show 87% kākā recall at detection, while species classification reaches 86% precision on kākā, with most errors occurring in low light, under strong occlusion, or with motion blur. 

AI-Driven Personal Health Baselines for Disaster Victim Rescue Prioritisation: A Non-Clinical Signal Approach 
Nishika Jayasinghe, Otago Polytechnic Auckland International Campus (presented by Sonia Gul)

Natural or man-made disasters result in people being trapped or displaced. This results in a veil of ignorance for rescue teams as they are unable to gain knowledge regarding the victims. This, in turn, creates a situation in which rescue teams need to determine whom to save first and who requires the most urgent attention. Recent advances in smartphones and wearable devices, such as sound detection and Bluetooth Low Energy (BLE) tracking, have made it easier to locate people trapped under debris. However, these methods do not distinguish between victims based on urgency. In these cases, rescue is usually, based on who is found first or who is closer to the rescue teams, rather than who needs help more urgently. This creates an important problem to address and a gap: current technologies can identify where a victim is, but they do not determine who should be rescued first when rescue teams have limited time and resources.

This study addresses the aforementioned problem by answering two research questions. RQ1: Can AI, trained on a person's pre-disaster, consented smartphone and wearable data, detect meaningful deviations from their individual baseline during a disaster when clinical assessment is unavailable? RQ2: Can AI-detected deviations be converted into a real-time priority score that improves rescue sequencing compared to prioritizing by detection order or proximity alone

The proposed method uses AI to learn each person’s normal health and activity pattern from their pre-disaster data, collected with their prior informed consent. During a disaster, the system can identify unusual changes in these patterns. It then combines this information with other factors, such as how long the person has been unable to communicate and the level of danger in their location, to create a clear rescue priority score. The system would use personal health data only with the person’s consent. Data sharing is deactivated by default and only activates during a declared emergency, at which point it is limited to rescue and medical response teams, helping address the privacy and ethical concerns of using personal health data.

By combining AI-based health risk information with existing victim-location technologies, this research aims to move disaster response from simply asking “Who can we find?” to “Who should we rescue first?” The proposed system could help search-and-rescue teams prioritize victims when direct medical assessment is not available.

Keywords: disaster response, wearable devices, personalized baseline modelling, anomaly detection, AI-driven triage, search and rescue, priority scoring, hidden evacuees

Faithful Evaluation of Semantic-ID Tokenizers for Generative Recommendation 
Qian Zhang, University of Otago 

Generative recommendation based on Semantic IDs (SIDs) represents each item as a discrete sequence of SIDs and is conventionally evaluated by matching the generated SID sequence against the target item's SID sequence. This evaluation is faithful only when each SID sequence uniquely identifies one item. In practice, SID collisions violate this condition: across the evaluated SID tokenizers and datasets, collision rates reach 30.52%, and SID-level Hit@10 is inflated by up to 103.36% relative to item-level Hit@10. To address this evaluation gap, we introduce Collision-Corrected Evaluation (CCE), which defines collision-aware item-level metrics (ItemHit@K, ItemNDCG@K) computed from generated SID sequences, and Zero-Collision Reassignment (ZCR), which constructs zero-collision SID assignments for existing tokenizers via minimum-cost reassignment. Applying these methods to four datasets and five representative SID tokenizers, we find that metric inflation scales with collision rate and is large enough to flip pairwise tokenizer comparisons under item-level re-evaluation. This finding calls into question the reliability of SID-level rankings reported in prior work and indicates that faithful tokenizer evaluation requires item-level correction or zero-collision SID assignments. 

TEMPO: Plug and play benchmarking for time-series ML pipelines 
Scott Mouat, University of Auckland 

Applying machine learning to time-series sensor data requires a coordinated sequence of windowing, feature extraction, feature selection, and predictive modelling. In practice, popular open-source libraries such as tsfresh, TSFEL, and catch22 operate with incompatible tabular schemas, rarely report runtime or memory demands, and lack unified experimental controls. This fragmentation prevents fair, reproducible benchmarking of interchangeable pipeline components across diverse physical domains and predictive tasks.

This project presents TEMPO (Time-series Evaluation for Model Performance Optimisation), a modular benchmarking framework developed to decouple and evaluate each stage of the time-series pipeline under controlled conditions. At its foundation is a standardised data contract paired with Apache Parquet storage, allowing datasets, extractors, selectors, and estimators to integrate through modular adapters. TEMPO unifies classification, regression, inverse regression, and forecasting workflows, incorporating on-disk caching, repeated cross-validation, post-hoc statistical hypothesis testing, and an asynchronous telemetry monitor that tracks CPU utilisation and process-tree Resident Set Size (RSS).

Evaluating real-world and synthetic pipelines exposed both computational bottlenecks and methodological limitations. On the computational side, deterministic caching reduced repeat feature retrieval to under 2 ms (a >2,700× speedup). Direct telemetry identified an 1,829 MB physical memory peak in tsfresh caused by internal DataFrame reshaping; native array and Numba-accelerated extractors reduced this footprint to approximately 400 MB. Methodologically, post-hoc statistical testing demonstrated that feature extraction libraries frequently achieve equivalent predictive accuracy across classification and forecasting tasks, whereas exogenous regression remained difficult across all tested methods.

By co-evaluating physical computational expenditure alongside predictive accuracy under strict leakage and statistical controls, TEMPO provides the foundation needed to transition time-series pipeline selection from ad-hoc experimentation into rigorous, reproducible engineering.

An AI-Enabled Multi-Agent Framework for Automating Student Assessment Evaluation 
Senaka Amarakeerthi, Otago Polytechnic AIC/Futureskills 

Manual evaluation of students' assessments places a substantial and repetitive workload on lecturers, particularly in short-cycle academic structures where timely feedback is critical to student learning. While large and small language models offer clear potential to support this process, institutional adoption is often constrained by concerns over student data privacy, data security, and compliance with institutional data governance policies when using public AI services. This paper proposes a generalisable AI-enabled multi-agent framework for automating assessment evaluation across diverse course types, designed to operate entirely within institutional infrastructure to preserve data privacy and security. The framework integrates agentic retrieval-augmented generation (RAG), semantic search and collaborative multi-agent reasoning within a modular architecture comprising four specialised agents: a coordinator agent, a retrieval agent, a summariser agent and a reasoning agent. These agents operate over a shared large language model (LLM) core layer, which provides a unified interface for agent interaction and supports the generation of automated evaluations and assessment reports. This design enables the framework to adapt to a range of assessment formats without modification to the underlying architecture. A human-in-the-loop feedback mechanism allows lecturer review and moderation decisions to be captured and used to incrementally refine the system's alignment with institutional marking practices over successive teaching blocks. As an initial proof of concept, the framework is being implemented on institutionally hosted infrastructure, with programming assessments used as the first case study for model training and evaluation using anonymised student assessment data. This paper presents the framework's design, architecture, and implementation approach for the programming assessment case study, and discusses its intended extension to other assessment types such as reports and project documentation. This work is expected to inform the framework's generalisability and provide a foundation for privacy-preserving, continuously improving assessment automation across broader educational contexts. 

Can AI Reduce Rural Health Inequities? Lessons from Healthcare Interventions and Policy Analysis in New Zealand, Australia and Canada 
Sharon Sitters, Manukau Institute of Technology & Unitec 

Background: Artificial intelligence (AI) is increasingly promoted as a solution to healthcare workforce shortages, service access barriers, and growing demand for healthcare services. For rural and remote communities, AI has the potential to improve healthcare access through enhanced screening, clinical decision support, and digital models of care. However, lop-sided implementation can perpetuate existing inequities in low-resource settings where limited infrastructure, and unequal access to resources are not resolved. This study uses evidence from AI healthcare interventions, to explore this gap between promise and practice for rural health settings, and the policy approaches that could help to promote health equity.

Methods: A scoping review of AI-enabled healthcare interventions in rural and remote populations was undertaken using the Health Equity Implementation Framework (HEIF). Literature published between 2015 and June 2025 was searched across MEDLINE, EMBASE, and CINAHL. Findings from the review informed a comparative policy analysis of national AI and digital health strategies from New Zealand, Australia, and Canada. Policy documents were analysed using the HEIF and the World Health Organization Digital Determinants of Health framework.

Results: Ten studies met inclusion criteria. AI interventions showed potential to improve access to screening, decision support, and chronic disease management in underserved communities. However, consistent implementation barriers included inadequate digital infrastructure, workforce shortages, limited integration and interoperability, concerns regarding data representativeness, amongst others. Comparative policy analysis revealed strong rhetorical commitments to health equity across all three countries but limited operationalisation of equity goals. Notably, policy gaps included insufficient attention to many of the challenges raised in the scoping review.

Conclusion: Although AI health policy frameworks in Aotearoa New Zealand, Australia, and Canada vary substantially, they are not yet sufficiently developed to support equitable AI implementation in rural and remote healthcare settings. Achieving health equity will require greater coordination between AI and digital health strategies, alongside targeted investment in digital infrastructure and workforce capability. Critically, governance structures must meaningfully embed the voices, priorities, and knowledge of rural, remote, and Indigenous communities throughout policy development and implementation processes. 

Evaluation of Retrieval-Augmented Generation for a university information chatbot 
Shreyas Patil, Southern Institute of Technology 

Artificial intelligence chatbots are able to provide prospective students with fast and easily accessible information about the university. However, general-purpose language models can sometimes provide inaccurate, unsupported responses or incomplete information that may mislead prospective students. Such responses would make the chatbot unreliable for institute-specific enquiries.

My research examines whether Retrieval-Augmented Generation (RAG) will improve the reliability of an AI chatbot that is designed to support prospective students during Southern Institute of Technology (SIT) Open Day events. A verified set of questions and answers is developed using official SIT programme information available on the official website. Multiple language models are initially evaluated using different prompting approaches to analyse the quality and reliability of the responses they generate. Based on these evaluations, the best-performing language model is selected for further RAG enhancement.

The selected model will be enhanced using RAG, which will allow it to retrieve relevant information from verified SIT sources and base its response on the retrieved information. The enhanced chatbot will then be compared with the original model using the same set of questions, answers, and evaluation process. This research aimed to determine whether grounding the AI chatbot in a verified information source improves response quality, reduces hallucinations and is a reliable source of information service for prospective students a more reliable information service for prospective students. 

Evaluating RAG for Reliable AI Chatbots in Educational Open Day Support 
Shreyas Patil, Southern Institute of Technology 

Artificial intelligence (AI) chatbots can provide prospective students with fast and easily accessible information about the programmes and services of educational institutes. Open Day events are particularly important because prospective students often require information about programmes, entry requirements, careers and study options. Providing this support usually requires staff from difference faculties and administration to be available to answer enquiries, creating a significant demand on institutional staff time and resources. An AI chatbot capable of providing reliable institution-specific information could help support these interactions. However, general-purpose Large Language Models (LLMs) have a tendency to generate inaccurate, incomplete, or unsupported responses, particularly when answering institute-specific questions. This study investigates whether Retrieval-Augmented Generation (RAG) can improve the reliability of an AI chatbot designed to support prospective students during Southern institute of Technology (SIT) Open Day Events.

A verified set of questions and answers is developed with the help of faculties to evaluate responses to realistic prospective-student enquiries. Multiple LLMs and prompting approaches are initially evaluated against the benchmark using pre-defined criteria for response quality and reliability. The highest performing model is then selected as the baseline model for the RAG enhancements. The enhanced model will allow to retrieve relevant information from verified SIT sources and base its response on the retrieved information.

The baseline and RAG-enhanced models would be evaluated using the same questions, reference answers, and evaluation process. The study focuses on whether grounding responses in verified institutional information improves factual accuracy and completeness while reducing unsupported or hallucinated information. The research aims to provide practical evidence on the effectiveness of RAG for reliable, institution-specific student information services and identify potential limitations of using LLM-based chatbots in educational settings.

Active Learning for Multi-output Regression 
Siyu Jin, University of Canterbury 

Active learning (AL) seeks to reduce labelling costs by choosing which observations to acquire for training. Many applications of machine learning involve tasks that require predictions of several quantities, making multi-output regression an important setting for AL. In this setting, a single AL query can inform predictions for multiple outputs, but its benefits may differ across them. Relationships between outputs offer information for assessing these benefits, yet those relationships must themselves be learned from limited data. Using these relationships to improve joint prediction does not establish how reliably they guide the selection of new observations. This research therefore aims to leverage relationships between outputs to make reliable AL acquisition decisions that prioritise greater reductions in prediction risk.

To address this challenge, this study examines when a learner should change its reliance on different acquisition criteria as learning progresses, and whether sensitivity to small changes in the labelled data provides useful evidence for these adjustments. The project focuses on multi-output batch selection, where decisions must account for both redundancy among selected inputs and information shared across outputs. Preliminary literature review work has informed a proposed analytical framework and taxonomy separating acquisition criteria from the learner's control of their use, to support comparison of existing strategies and investigation of new combinations. Insights from these investigations will guide adaptive strategies that switch between acquisition criteria or adjust their weights in multi-output regression. 

Adaptive Governance for AI-Enabled Disaster Infrastructure: Resilience–Trust Operating Boundaries Under Infrastructure Degradation 
Sonia Gul, Otago Polytechnic - Auckland International Campus (OPAIC) and FutureSkills 

Artificial intelligence is increasingly embedded within disaster response infrastructure, supporting prediction, coordination, and resource allocation across distributed sensor, edge, and cloud systems. A persistent limitation in this space is the treatment of governance: mechanisms such as access control, auditing, and accountability protocols are typically designed for stable operating conditions and implemented as static, externally imposed rules. Under disaster conditions, where connectivity is intermittent and infrastructure is fragmented, such mechanisms frequently become inconsistent, delayed, or unenforceable, precisely when reliable and accountable decision-making matters most. Post-event analyses of recent disaster responses, including Cyclone Gabrielle, further indicate that centralised governance structures can struggle to adapt as conditions degrade, weakening both coordination effectiveness and legitimacy.

This work proposes an Adaptive Governance Framework for AI-enabled disaster infrastructure, reconceptualising governance as a resilience-aware operational layer embedded within system architecture rather than an external policy overlay. The framework integrates a layered Field, Edge, and Cloud architecture with an AI-assisted governance orchestration mechanism that monitors infrastructure conditions and recommends governance mode transitions, while final authority to enact those transitions remains with human decision-makers. A key contribution is the definition of resilience–trust operating boundaries, the conditions under which governance mechanisms remain both technically enforceable and defensible as infrastructure degrades, reframing governance as a context-sensitive capability that scales with system state rather than a fixed or binary property.

The framework structures this scaling through four graduated governance modes, triggered by the degree of infrastructure and connectivity loss: centralised oversight under stable conditions, distributed authority across regional agencies under partial disruption, community-assisted governance recognising local and iwi leadership when official channels are severely degraded, and emergency governance for life-safety decisions under near-total disruption. This conceptual framework will be illustrated through a planned scenario-based analysis of a flood-response context. This work complements the Catalyst-funded MHEAS project (Mobile and Home Evacuee Assistance System), presented at this conference in 2025, by addressing the governance dimension of AI-enabled disaster infrastructure alongside MHEAS's technical focus on federated, privacy-preserving evacuee support.

Víðarr–Norn: Menger-KAN with Tree Numbers, a Finite Form of Kolmogorov’s 1956 Representation 
Todd Cochrane, Nelson Marlborough Institute of Technology 

Víðarr–Norn tests Menger-KAN, a tree-valued Kolmogorov–Arnold network, for learning and computation. Kolmogorov's 1956 theorem writes every continuous function of n variables as a sum of n + 1 functions of tree-valued inner maps. Their values lie in Menger's universal tree, a loop-free continuum containing a copy of every dendrite. Menger-KAN makes this representation finite and trainable. Each input becomes n + 1 tree numbers. A tree number is a point of a searched finite tree: digits name its branch, and a fraction its position. As in the theorem, the inner maps depend on the inputs alone. The outer functions that read the tree numbers are fitted exactly by least squares. Across four regression tasks with eight inputs each, this form beat linear regression in all 16 cases, at half its average test error. Its error was still about 34 times that of a parameter-matched multilayer perceptron (MLP). Training the inner maps on the target multiplied this error by 0.42. Where level sets branch, a forest of different trees beat every MLP tested. On a double-well function, a code on its level-set tree explained all the test variance; a single-branch code explained none. For 2 to 32 inputs, more terms stopped helping at 1.5 to 2.2 times n + 1. This is nearer the 1957 theorem's 2n + 1 than n + 1. Kolmogorov's division of labour therefore works in finite form. Trees help where level sets branch. On the other three tasks, the matched MLP beat every tree-based model. 

TEMPO: Plug and play benchmarking for time-series ML pipelines 
Trisan Ethan Q Huynh, University of Auckalnd 

Applying machine learning to time-series sensor data requires a coordinated sequence of windowing, feature extraction, feature selection, and predictive modelling. In practice, popular open-source libraries such as tsfresh, TSFEL, and catch22 operate with incompatible tabular schemas, rarely report runtime or memory demands, and lack unified experimental controls. This fragmentation prevents fair, reproducible benchmarking of interchangeable pipeline components across diverse physical domains and predictive tasks.

This project presents TEMPO (Time-series Evaluation for Model Performance Optimisation), a modular benchmarking framework developed to decouple and evaluate each stage of the time-series pipeline under controlled conditions. At its foundation is a standardised data contract paired with Apache Parquet storage, allowing datasets, extractors, selectors, and estimators to integrate through modular adapters. TEMPO unifies classification, regression, inverse regression, and forecasting workflows, incorporating on-disk caching, repeated cross-validation, post-hoc statistical hypothesis testing, and an asynchronous telemetry monitor that tracks CPU utilisation and process-tree Resident Set Size (RSS).

Evaluating real-world and synthetic pipelines exposed both computational bottlenecks and methodological limitations. On the computational side, deterministic caching reduced repeat feature retrieval to under 2 ms (a >2,700× speedup). Direct telemetry identified an 1,829 MB physical memory peak in tsfresh caused by internal DataFrame reshaping; native array and Numba-accelerated extractors reduced this footprint to approximately 400 MB. Methodologically, post-hoc statistical testing demonstrated that feature extraction libraries frequently achieve equivalent predictive accuracy across classification and forecasting tasks, whereas exogenous regression remained difficult across all tested methods.

By co-evaluating physical computational expenditure alongside predictive accuracy under strict leakage and statistical controls, TEMPO provides the foundation needed to transition time-series pipeline selection from ad-hoc experimentation into rigorous, reproducible engineering. 

PACEAttention: Principled and Adaptive Feature Compression-Expansion Grounded in the Geometry of MCR2 
Xiaojie Yu, University of Otago 

The maximal coding rate reduction (MCR2) objective is proposed for learning low-dimensional subspace representations and for principled deep model design, where layer structures are derived by unrolling its optimization steps.However, existing methods motivated by this objective do not fully adhere to design principles implied by the MCR2 gradient, which weakens the principled and interpretable foundations of the resulting models. In this work, we introduce PACEAttention, a novel principled attention mechanism inspired by the textit{geometric insight }of MCR2, whose gradient-based updates move features along directions shaped by the underlying low-dimensional feature structure. Our method captures this structure by leveraging randomization to guide feature updates. This principled construction enables the resulting PACENet to exhibit enhanced interpretability, with different heads attending to distinct image regions and capturing textit{fine-grained} structures under simple supervised training.Experiments demonstrate that PACEAttention achieves superior performance and more stable scalability than previous principled modules while remaining low complexity.