Output diversity and voting match neural thicket results in ai models
Rethinking Neural Thickets Without Weight Perturbations
Machine LearningArtificial IntelligenceComputer Vision and Pattern Recognition
Summary
Large AI models have been thought to hold many expert versions close together inside their 'brain' weights, discoverable by adding random noise. The authors show that you don’t need to change all the model’s weights to get better answers; just tweaking or voting on outputs works as well or better. They explain this by creating many slightly different outputs, picking the best ones based on a small test, and then combining their answers to improve performance. This approach works across language and image tasks without changing the model’s internals or retraining it.
What this means in practice
- •For machine learning engineers: Improve model performance on reasoning and image tasks by ensemble voting over output variations without retraining or changing model weights.
- •For earth observation analysts: Increase robustness of vision models under domain changes using output-level voting instead of full model perturbations.
Authors
Nicolas Dufour, Elouan Gardès, Amil Dravid, Patrick Pérez, Loic Landrieu
Abstract
The Neural Thickets hypothesis posits that pretrained language models lie in dense neighborhoods of task-specific experts in weight space, which random Gaussian perturbations can uncover. We reproduce the reported gains, but show that they do not require this geometric explanation. Perturbing a single decoder block already recovers much of the improvement, while keeping the model entirely frozen and adding a random bias to its output logits matches or exceeds full-weight perturbations on most reasoning tasks. This suggests an alternative explanation based on output diversity, validation-based selection, and voting. We formalize this view as Perturb-Select-Vote (PSV), a general procedure that generates perturbed predictors, selects high-performing members on a small validation set, and aggregates their predictions. This interpretation connects Neural Thickets to established inference-time methods, including sampling, self-consistency, and verifier-guided selection. A first-order analysis further shows that small weight perturbations act primarily through their induced logit shifts. We extend this analysis to vision, where logit PSV substantially improves performance under domain shift, including on Earth observation, while full-weight perturbations provide little additional benefit. Together, these results suggest that the gains attributed to nearby weight-space experts can instead be obtained through output diversity, selection, and voting. Logit PSV realizes these benefits with a fully frozen model, providing a simple, training-free alternative to full-weight perturbation ensembles across language and vision.