Accessibility Plasticity: A New Principle for Adaptive Computation in Neural Networks

A new arXiv preprint (2607.22748) introduces Accessibility Plasticity, a principle of adaptive computation for neural networks. The authors argue that current AI systems primarily adapt by modifying parameters within fixed computational structures, overlooking the distinction between computational capability and computational accessibility. They propose that systems should also adapt by reorganizing which existing computations can interact. The work formalizes this through a relationship-based operational realization and establishes a reuse-first hierarchy, where accessibility changes precede more costly capability or structural modifications. A proof-of-concept on sequential learning tasks shows that accessibility adaptation reduces the need for capability changes while maintaining performance. The paper is categorized as a cross-type announcement and has not yet been peer-reviewed.