Episode

"Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations" by Subhash Kantamneni, kitft, Euan Ong, Sam Marks

Podcast
LessWrong (Curated & Popular)
Published
May 8, 2026
Duration seconds
1092
Processing state
not_requested
Canonical source
https://www.buzzsprout.com/2037297/episodes/19144572-natural-language-autoencoders-produce-unsupervised-explanations-of-llm-activations-by-subhash-kantamneni-kitft-euan-ong-sam-marks.mp3
Audio
https://www.buzzsprout.com/2037297/episodes/19144572-natural-language-autoencoders-produce-unsupervised-explanations-of-llm-activations-by-subhash-kantamneni-kitft-euan-ong-sam-marks.mp3
JSON
/v1/public/podcasts/lesswrong-curated-popular-5643401/episodes/natural-language-autoencoders-produce-unsupervised-explanations-of-llm-activations-by-subhash-kantamneni-kitft-euan-ong-sam-marks
Markdown
/podcast/lesswrong-curated-popular-5643401/natural-language-autoencoders-produce-unsupervised-explanations-of-llm-activations-by-subhash-kantamneni-kitft-euan-ong-sam-marks.md

Actions

  • POST https://stenobird.com/v1/public/podcasts/lesswrong-curated-popular-5643401/episodes/natural-language-autoencoders-produce-unsupervised-explanations-of-llm-activations-by-subhash-kantamneni-kitft-euan-ong-sam-marks/transcription-requests
    Idempotently request low-priority transcript generation for this episode.
  • GET https://stenobird.com/podcast/lesswrong-curated-popular-5643401/natural-language-autoencoders-produce-unsupervised-explanations-of-llm-activations-by-subhash-kantamneni-kitft-euan-ong-sam-marks.md
    Read the agent-friendly Markdown representation of this episode resource.

Summary

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 activation reconstructor (AR) that maps the description back to an activation. We jointly train the AV and AR with reinforcement learning to reconstruct residual stream activations. Although we optimize for activation reconstru...