Welcome to the NeurIPS 2026 Workshop Closed-Loop NeuroAI: Scalable Biological Priors for Adaptive Intelligence! Building on the success of our inaugural event, this second NeuroAI workshop will take place at NeurIPS 2026 in Sydney and will continue to bring together researchers and practitioners from neuroscience, artificial intelligence, cognitive science, neurotechnology, and related fields.

We are in a period of rapid progress in artificial intelligence, marked by the rise of large-scale models in language, vision, multimodal learning, generative modeling, and autonomous agents. As these systems scale, there is growing interest in understanding which principles of biological intelligence can help AI systems become more robust, efficient, adaptive, interpretable, and aligned with real-world constraints. The field of NeuroAI – at the intersection of artificial and natural intelligence – aims to bridge neuroscience and AI by translating insights from brains, bodies, and behavior into computational principles for modern machine learning.

This year’s workshop focuses on a central question: which biological priors scale, and how can they support adaptive intelligence in closed-loop, real-world settings? We are particularly interested in work that goes beyond analogy and evaluates whether neuroscience-derived mechanisms measurably improve learning, generalization, adaptation, robustness, or interpretability under realistic scaling regimes.

The workshop will be organized around the following themes:

• Priors that scale: inductive biases from brains that improve modern learning.
  At the center of this workshop is a simple question: which neuroscience-derived priors measurably help contemporary AI systems under realistic scaling regimes? We invite work that operationalizes biological constraints and biases, including modularity, hierarchy, recurrence, sparsity, energy and latency constraints, structured memory and control, and action-conditioned computation. We are especially interested in studies that evaluate their impact on sample efficiency, robustness, adaptation, and out-of-distribution generalization in modern architectures such as Transformers, diffusion models, state-space models, and agents. Submissions that include explicit comparisons against strong scaling baselines, cross-scale evaluations, or careful reports of failure modes are particularly encouraged.

• Learning rules and credit assignment: from plasticity to scalable optimization.
  A long-standing promise of NeuroAI is that brains learn effectively from limited supervision, sparse feedback, and continual interaction. This theme focuses on biologically grounded learning rules and training signals, including local plasticity rules, surrogate gradients, three-factor learning, predictive learning, and approximate credit assignment. We welcome work that asks when these mechanisms provide practical advantages in modern machine learning settings such as online learning, continual learning, low-power edge deployment, and adaptive agents. We also encourage mechanistic and theoretical contributions that clarify when such rules approximate, complement, or challenge backpropagation at scale.

• Self-supervised prediction and active perception: scalable predictive principles.
  Brains learn by predicting and acting, not simply by consuming labeled datasets. This theme covers self-supervised learning principles inspired by predictive coding and active inference, including predictive representation learning, world-model learning, uncertainty-aware objectives, and closed-loop action-perception learning. We encourage submissions that connect these principles to scalable training pipelines and evaluate them on challenging transfer settings. We are also interested in embodied and hybrid bio-silico platforms that stress-test predictive learning under real-time constraints. A key question is whether predictive and self-supervised principles scale in neural time-series settings, and whether progress comes primarily from larger cross-subject datasets, stronger inductive biases, or subject-specific and closed-loop adaptation.

• Evaluation beyond decoding: benchmarks, neural alignment, and “what fails to transfer”.
  As brain foundation models, neural decoding systems, and biologically inspired agents proliferate, NeuroAI urgently needs evaluation practices that test more than downstream accuracy. This theme focuses on benchmarks and mechanistic evaluation protocols for asking whether proposed neural priors genuinely transfer across subjects, tasks, modalities, scales, and deployment settings. We welcome work on standardized evaluation for brain and body foundation models, neural and behavioral alignment metrics, causal interventions, circuit-level analyses, and tests of whether model representations are brain-like beyond surface-level performance. We explicitly encourage careful negative results, ablations, and failure-mode analyses, including cases where a neuro-inspired idea does not help under scale, fails to generalize across datasets, or collapses to a shortcut. This also includes work comparing representations across scales and modalities, for example from population imaging to invasive recordings or from one measurement window to another, and asking what structure is preserved, lost, or aliased under downsampling, intervention, or modality shifts.

• Closed-loop neuroadaptive systems: co-adaptation, BCIs, and real-world constraints.
  Transfer is ultimately judged in deployment: under drift, non-stationarity, limited calibration time, and human-in-the-loop feedback. This theme covers closed-loop neuroadaptive interfaces and neurotechnology, including brain-computer interfaces, co-adaptive decoding, adaptive stimulation, and hybrid systems that integrate language models with neural signals for assistive communication and cognition. We particularly encourage work that treats closed-loop interaction as a scaling test for priors: whether a proposed mechanism yields stable learning and robust performance under real-time constraints and continual adaptation. We also welcome work probing whether high-dimensional and potentially compositional neural representations support flexible closed-loop behavior, and whether these mechanisms suggest alternatives to current end-to-end neural architectures.

Our workshop aims to explore the intersections among these research areas and provide a platform for researchers to navigate the links between artificial and natural intelligence. We hope to use the workshop as a vehicle for identifying which biological principles can meaningfully scale, which fail to transfer, and which require new forms of evaluation, interaction, and theory. Through invited talks, contributed papers, posters, and discussions, we seek to clarify fundamental gaps and challenges in NeuroAI and to help shape future directions for adaptive, robust, and biologically grounded intelligence.

For more details, please see the Call for Papers.

For further information and should you have any inquiries, please contact: neuroai.neurips2024@gmail.com