AI · Studies · 2026-10-30 08:43
The paper finds that large language models can recognize dropout and Gaussian noise applied to their activations, offering a new perspective on internal state awareness and robustness in LLMs.
Auto-generated summary for reference only. RSS