Tag
The paper studies which token positions in natural language autoencoders are most informative to explain, finding that chat-structure-based ranking selects relevant explanations better than computational signals, and that explaining only 5% of positions retains nearly all audit success on prompt injection and concealment tasks.
This pilot study uses Natural Language Autoencoders to probe whether Qwen2.5-7B internally represents Colombian identity, socioeconomic status, or stereotype-related information when processing Colombian-Spanish and English prompts, finding evidence of latent inferences before they are verbalized.
The author built a custom llama.cpp server and Mikupad UI to enable local inference and activation steering with Anthropic's open-weight Natural Language Autoencoders. A LoRA version is in development to reduce memory requirements.
Anthropic and Neuronpedia released research and tools on Natural Language Autoencoders (NLA), enabling users to view the internal 'thoughts' of Gemma 3 during token generation. The release includes model weights for the Auto Verbalizer and Activation Reconstructor, hosted on Hugging Face and Neuronpedia.
Anthropic introduces Natural Language Autoencoders (NLAs), a method to translate internal AI activations into human-readable text, enabling better understanding of model thoughts and improving safety by revealing hidden reasoning processes.
Anthropic and Neuronpedia have partnered to release Natural Language Autoencoders (NLAs) on open models, allowing researchers to gain hands-on experience with this interpretability tool.