LoRA sentiment behavior collapses under Gaussian weight noise at about 3.5x adapter RMS

A researcher probed how robust a fine-tuned sentiment adapter on a 4-bit Qwen 3B model is by adding Gaussian noise scaled to multiples of the adapter's RMS to every LoRA tensor. Accuracy on 400 SST-2 test rows stayed near 95% up to roughly 2x RMS, then dropped sharply, with the half-gain point at about 3.5x RMS. At 5x and above, accuracy fell to chance level, below the 60.8% base-model score, meaning the noise destroyed the adapter's contribution rather than restoring the base behavior. Quantization noise was found to sit well inside the stable plateau. The author cautions that one adapter, one task, and a single seed per noise level are insufficient to establish a general pattern, and outlines follow-up tests across other tasks and adapter components.