Post-training a language model for a specific task usually requires substantial data and compute. A recent paper suggested a shortcut: simply adding random noise to the model's weights could uncover hidden task experts. This would mean you could get specialized abilities without any additional training.
We tested this claim by examining whether the apparent expert performance came from genuine skill or just lucky variation on the selection questions. We found that for most tasks, more than 75% of the top expert's advantage could be explained by chance alone. As we used more selection questions, this advantage shrank quickly.
On new, never-before-seen questions, the top expert’s advantage nearly vanished, often within the normal variability of the original model. We found that combining several perturbed models gave no better results than simply using multiple outputs from the original model. This shows that random weight perturbations do not reliably find task experts.