LLMs recognize dropout and Gaussian noise applied to their activations

A paper posted to arXiv on 19 April 2026 by Damiano Fornasiere reports that eight language models from the Gemma, Llama, Mistral, Olmo, Qwen and Seed families, sized between 8B and 36B parameters, can detect and localize dropout or Gaussian noise applied to their activations at a target sentence, often with perfect accuracy on multiple-choice questions such as "Which of the previous sentences was perturbed?" The study finds Olmo models develop the capacity during pre-training and rules out lower log-likelihood of perturbed sentences as the explanation, while zero-shot accuracy at identifying which perturbation was applied correlates with perturbation strength for Qwen3-32B and Qwen3.8-27B. The authors discuss a possible data-agnostic "training awareness" signal and its implications for AI safety, noting dropout is used in training regularization while Gaussian noise is sometimes added at inference.

Computer Science Artificial Intelligence Submitted on 19 Apr 2026 v1 https://arxiv.org/abs/2604.17465v1 , last revised 14 Oct 2026 this version, v3 Title:Language models recognize dropout and Gaussian noise applied to their activations View PDF https://arxiv.org/pdf/2604.17465 Abstract:We provide evidence that language models can detect, localize and, to a degree, verbalize the difference between a-semantic perturbations applied to their activations. More precisely, we either a apply dropout or b add Gaussian noise to them, at a target sentence. We then ask a multiple-choice question such as "Which of the previous sentences was perturbed?" or "Which perturbation was applied?". We test eight models from the Gemma, Llama, Mistral, Olmo, Qwen and Seed families, with sizes between 8B and 36B, all of which can easily detect and localize the perturbations, often with perfect accuracy. To investigate the mechanisms, we find that Olmo models develop this capacity during pre-training, and we rule out the hypothesis that perturbed sentences simply have lower log-likelihood. Notably, the zero-shot accuracy of Qwen3-32B and Qwen3.8-27B in identifying which perturbation was applied correlates with the perturbation strength. While this task proves hard for the remaining models, most of them can learn, when taught in context, to distinguish between dropout and Gaussian noise. Interestingly, Qwen3-32B and Qwen3.8-27B learn worse when supervised with wrong labels, again suggesting a prior towards the correct ones. Because dropout is used as a training-regularization technique, while Gaussian noise is sometimes added during inference, we discuss the possibility of a data-agnostic "training awareness" signal and the implications for AI safety. Submission history From: Damiano Fornasiere view email https://arxiv.org/show-email/0b81ff33/2604.17465 Sun, 19 Apr 2026 14:30:13 UTC 94 KB \ v1\ https://arxiv.org/abs/2604.17465v1 Thu, 30 Apr 2026 22:17:02 UTC 94 KB \ v2\ https://arxiv.org/abs/2604.17465v2 v3 Wed, 14 Oct 2026 19:43:24 UTC 6,905 KB Additional Features References & Citations Loading... Bibliographic and Citation Tools Bibliographic Explorer What is the Explorer? https://info.arxiv.org/labs/showcase.html arxiv-bibliographic-explorer Connected Papers What is Connected Papers? https://www.connectedpapers.com/about Litmaps What is Litmaps? https://www.litmaps.co/ scite Smart Citations What are Smart Citations? https://www.scite.ai/ Code, Data and Media Associated with this Article alphaXiv What is alphaXiv? https://alphaxiv.org/ CatalyzeX Code Finder for Papers What is CatalyzeX? https://www.catalyzex.com DagsHub What is DagsHub? https://dagshub.com/ Gotit.pub What is GotitPub? http://gotit.pub/faq Hugging Face What is Huggingface? https://huggingface.co/huggingface ScienceCast What is ScienceCast? https://sciencecast.org/welcome Demos Recommenders and Search Tools Influence Flower What are Influence Flowers? https://influencemap.cmlab.dev/ CORE Recommender What is CORE? https://core.ac.uk/services/recommender arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs https://info.arxiv.org/labs/index.html .