Neural Thickets
A theoretical hypothesis about the weight space of pretrained models, positing that expert models specialized for different tasks are densely distributed in the neighborhood of a pretrained model's weights and can be activated through random perturbations.
Timeline (last 90 days)
Gan & Isola (2026) propose the neural thicket hypothesis: a pretrained model lies in a dense set of task experts, and simple Gaussian weight perturbations can reveal these experts.
A recent study re-examining neural thickets found that in 10 of 12 model-task combinations, more than 75% of the top-expert reward improvement could be predicted by a null model based solely on selection noise.
The main disagreement between Gan & Isola's original work and the re-examination study lies in whether the improvement from perturbations generalizes and in the effectiveness of the majority-voting strategy.
All Facts (3)
Gan & Isola (2026) propose the neural thicket hypothesis: a pretrained model lies in a dense set of task experts, and simple Gaussian weight perturbations can reveal these experts.
50%UnverifiedA recent study re-examining neural thickets found that in 10 of 12 model-task combinations, more than 75% of the top-expert reward improvement could be predicted by a null model based solely on selection noise.
50%UnverifiedThe main disagreement between Gan & Isola's original work and the re-examination study lies in whether the improvement from perturbations generalizes and in the effectiveness of the majority-voting strategy.
50%