Amazing growth in Large Language Models (LLM) has shifted the state of the art away from whether narratives can be generated whilst simulating different perspectives to needing strong conceptual and artificial models of perception as well as designing appropriate evaluation frameworks for subjective plausible narratives. As Artificial Intelligence and its role in Human daily living become more intertwined than ever, the project has shifted to more general questions regarding the evaluation of narratives or explanations of internal cognitive states.
Given the above shift, the project has also needed to align better with anticipated potential applications (reflection support technology being one of the most relevant). This has been addressed through additional activities related to writing 2 position papers (under development) that try to address this issue, and strategically collaborating more with potential stakeholders of NEON. As well as investing more effort in laying ground work for potential use cases for the results of NEON. This has taken a multi-layer strategy. At the practical level, we have targeted specific families of potential downstream applications in health, education, enterprise, and industry with collaborators specifically specialised to those domains, identifying specific use cases, how the quality of intention narratives would be useful. At the meta level, what are the right research practices for the new era of LLM research that is closely intertwined with humans? To this end, I co-initiated a new community initiative to bring machine learning researchers together to shine a light on machine learning research practices and to ask what are pros and cons for different ways of carrying out of research (resulting in 3 workshops in 2025 and a planned podcast). In addition, I am co-editing an interdisciplinary handbook on group and team dynamics that discussed interdisciplinary work flows for future research between social sciences and computer science.
With this in the background, the first major achievement has been the design and collection of a dataset that records self-reported goals, participant context, and their desire to connect with others. We will use this to study the relationship between long term goals and perceived momentary intentions. Specifically, we organised a instrumented living lab workshop in July 2025 at a real conference. This scientific workshop had multiple goals. The first was to bring researchers from different disciplines together who were interested in the use of sensors and computational methods to analyse group dynamics. The event had 32 people and measured their social behaviour via microphones, cameras, and wearable sensors. This ethical board approved dataset will be shared with the research community under a restrictive end user license. To make this event possible, we had to augment the existing fully synchronised multimodal multisensor recording setup to allow for full 3d video reconstruction of people's body keypoints and also full speech quality audio. In addition, we developed a parallel data set from smart wearable sensors enabling investigations of more pervasive privacy sensitive design approaches for analysing social intentions. We adapted our ID badge casing into a drinking glass format that allowed for hand movements from the glass holding hand to be captured.
Other achievements include the first investigation into the detection of unrealized intentions to speak, which we initially thought would not be visible or perceptible by an external observer. It turns out that these behaviours are perceptible and carry behavior that can be detected to some degree.
We also investigated the efficacy of using pre-trained F-formation detectors for analyzing child group interactions. Unfortunately, this was not possible and led to open questions about whether children also form F-formations and from what age. These are relevant for the analysis of social intentions of children, which will are planned for later in the project.
To contextualise the evaluation of intention narratives, we carried out a survey on reflection support technology for social interactions. We found that the state of the art relies on subjective memories to reflect on social interactions. This calls for the need for narratives that are grounded in observable cues from what actually happened during the interaction narratives that can be used to argue and help a user to reflect on past social interactions.
