Questions over whether neural networks learn universal or model-specific representations framed a community event at the Cognitive Computational Neuroscience conference in August 2025, highlighting future directions on a fundamental topic in NeuroAI.

Powerful generative AI models for designing biological macromolecules are being developed, with applications in medicine, biotechnology and materials science, but these models are expensive to train and modify. Leyva et al. introduce the Key-Cutting Machine, an optimization-based platform for proteins and peptides that iteratively leverages structure prediction to match desired backbone geometries.

Graber et al. characterize biases and data leakage in protein–ligand datasets and show that a cleanly filtered training–test split leads to improved generalization in binding affinity prediction tasks.

ALL-conformations, a dataset capturing the full range of experimentally observed conformations of CDR loops, T cell and antibody regions interacting with antigen targets, is introduced. ITsFlexible—a deep learning tool trained on this new dataset—advances predictions of immune receptor structural dynamics.

A deep learning approach, AdaptiveNN, shifts machine vision models from passive to active to mimic human-like perception. The method achieves inference costs that are up to 28-times lower without accuracy loss, while showcasing online-adaptable and interpretable behaviours.

Xiao et al. introduce ‘capability density’, defined as capability per parameter, as a metric for evaluating large language models. They report an empirical trend, the ‘densing law’, which states that capability density doubles approximately every 3.5 months, indicating that equivalent model performance can be achieved with exponentially fewer parameters over time.

Suzgun et al. find that current large language models cannot reliably distinguish between belief, knowledge and fact, raising concerns for their use in healthcare, law and journalism, where such distinctions are critical.

GraphPep presents an interaction-derived and protein language model-powered graph learning framework for robust scoring of protein–peptide complexes, substantially enhancing the binding mode prediction of protein–peptide docking.

Powerful generative AI models for designing biological macromolecules are being developed, with applications in medicine, biotechnology and materials science, but these models are expensive to train and modify. Leyva et al. introduce the Key-Cutting Machine, an optimization-based platform for proteins and peptides that iteratively leverages structure prediction to match desired backbone geometries.

Graber et al. characterize biases and data leakage in protein–ligand datasets and show that a cleanly filtered training–test split leads to improved generalization in binding affinity prediction tasks.

Questions over whether neural networks learn universal or model-specific representations framed a community event at the Cognitive Computational Neuroscience conference in August 2025, highlighting future directions on a fundamental topic in NeuroAI.

Most policy proposals aimed at managing the risks of artificial intelligence (AI)-enabled weapons rely heavily on meaningful human control or appropriate human judgment for risk mitigation. This Comment argues that there are various ways humans can exert such control over AI, and that developing a careful taxonomy of these is necessary for building actionable risk-mitigation policies for warfighting AI.

The organizers reflect on how a multi-year, multi-country benchmark aligned AI research in road damage detection with practical and regional constraints, steering it towards deployment relevance.