Adding noise to frozen model outputs outperforms weight tweaking in inference and vision tasks
A hypothesis has long dominated AI research halls, assuming that pre-trained language models reside within dense engineering environments of specialist experts in weight space, and that extracting their latent capabilities requires injecting random Gaussian noise into the model’s weights. However, a new paper led by researcher Nikola Dufour with Ilwan Gard, Amal Drafid, Patrick Perez and Loïc Landrio re-examined this hypothesis, known as “neural amulets”, and found that the tangible performance gains do not require this complex engineering explanation nor do they call for altering the model’s original weights.
The researchers showed, when reproducing the experimental results, that injecting noise into a single decoder block of the model recovers a large portion of the performance improvement, whereas keeping the model fully frozen while adding a random bias to the logits in the outputs matches or even surpasses the gains from full-weight noise in most inference tasks. The first-order mathematical analysis presented by the team demonstrated that the slight noise applied to the weights essentially operates by shifting the logits produced by the model, shifting the explanation of the phenomenon from a hypothesis of neighboring experts in weight space to mechanisms of output diversity and verification-based selection.
The team framed this experimental pathway as a general procedure they called “Perturb, Select, Vote” (PSV).The methodology generates noisy predictions, then selects the best-performing ones based on a small validation set, and aggregates the final predictions through a voting mechanism. This analysis links the behavior of neural amulets to established inference methods such as sampling, self-consistency and verifier-guided selection, showing that the recorded gains do not require replicating heavy model copies in computational memory, but can be achieved directly through statistical control of inference outputs.
The team’s experiments extended to computer-vision models under domain-shift conditions, particularly in earth-monitoring and remote-sensing tasks.Applying noise to the logits achieved a substantive improvement in the accuracy of vision models when facing data-distribution shifts.Conversely, full-weight noise offered no notable additional benefit, confirming that processing probability outputs in a fully frozen model provides an effective, general alternative that obviates weight noise or extra training in both language and vision domains.
This computational shift imposes a direct operational reality on engineering teams and infrastructure developers in the Gulf, Egypt and the Levant. Instead of allocating massive video-memory capacity and exhausting GPUs to load duplicated copies of model weights for running prediction ensembles or locally tweaking parameters, the results allow system administrators in banks, government agencies and regional research centers to keep the source model in a fully frozen state and run the perturb-select protocol only on the logits outputs. This translates into a tangible reduction in iterative inference cost and enables the use of consistency and voting techniques without duplicating infrastructure. It also gives developers of land-monitoring and urban-planning applications in the region a straightforward tool to boost remote-sensing model accuracy in desert environments and local terrain changes without incurring training costs or altering the approved open models.
These results reveal that the shortest path to improving model reliability does not necessarily pass through the complexity of weight space or the pursuit of presumed engineering configurations within neural networks. Our examination of the experiment confirms that engineering the outputs at inference time gives the AI engineer efficiency that surpasses structural-noise methods, and shows that frozen models retain higher inferential capability than their raw weights suggest whenever statistical diversity is well extracted from their outputs.