Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations — LessWrong

Abstract > We introduce Natural Language Autoencoders (NLAs), an unsupervised method for generating natural language explanations of LLM activations.…

x Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations — LessWrong AI Frontpage 2026 Top Fifty: 37 % 213 Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations by Subhash Kantamneni , kitft , Euan Ong , Sam Marks 7th May 2026 AI Alignment Forum 10 min read 35 213 Ω 77 Abstract We introduce Natural Language Autoencoders (NLAs), an unsupervised method for generating natural language explanations of LLM activations. An NLA consists of two LLM modules: an activation verbalizer (AV) that maps an activation to a text description and an activa

saved by

• Kevin Zhang
• Anastasia Wei

related reading

• Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activationstransformer-circuits.pub
• Natural Language Autoencoders \ Anthropicanthropic.com
• Can activation verbalizers surface an internal chain of thought? — LessWronglesswrong.com
• GitHub - kitft/natural_language_autoencoders · GitHubgithub.com
• An Intuitive Explanation of Sparse Autoencoders for LLM Interpretability | Adam Karvonenadamkarvonen.github.io
• Mechanistically Eliciting Latent Behaviors in Language Models — AI Alignment Forumalignmentforum.org
• GitHub - asherps/EasyNLA: Minimal codebase for efficiently training Natural Language Autoencoders (NLAs). Built on Celeste's nanoNLA: https://github.com/ceselder/nanoNLAgithub.com
• Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers — LessWronglesswrong.com
• Language models can explain neurons in language modelsopenaipublic.blob.core.windows.net
• A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Modelsarxiv.org
• Prediction, Explanation, or Over-interpretation?elena-baixy.github.io
• [2511.08579] Training Language Models to Explain Their Own Computationsarxiv.org