"Shake" LLMs to make them better...?

New AI research called 'neural thickets' reveals that after pre-training, large language model weights don't settle at a single optimal solution but reside in a dense region surrounded by alternative specialist models. By slightly perturbing weights with Gaussian noise ('jiggling'), the model can be nudged toward task-specific specialists — improving math, coding, or writing capabilities. This effect only manifests in very large models, which have sufficient parameter space to contain these hidden specialists. The implication is that a pre-trained model is effectively a neighborhood of latent specialists, and random weight sampling with outcome-based evaluation can directly yield improved models.