Rethinking Neural Thickets Without Weight Perturbations

In 30 Sekunden

Was passiert ist

Researchers show that Neural Thickets gains come from output diversity and voting, not weight-space geometry. Frozen models with random logit perturbations match or exceed full weight perturbations.

Warum es zählt

Matters for engineers optimizing inference ensembles on language and vision tasks, especially under domain shift or reasoning workloads.

Achtung

Results are empirical on specific tasks and validation sets. Generalization to other model architectures, scales, or domains remains unclear.