--- title: SAEs trained on the same data don’t learn the same features description: Research from EleutherAI shows that TopK Sparse Autoencoders (SAEs) trained on identical data with the same batch order but different random seeds share only... canonical: https://daily.dev/posts/saes-trained-on-the-same-data-don-t-learn-the-same-features-2dsq5rc2m twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: SAEs trained on the same data don’t learn the same features | daily.dev og:description: Research from EleutherAI shows that TopK Sparse Autoencoders (SAEs) trained on identical data with the same batch order but different random seeds share only... og:url: https://daily.dev/posts/saes-trained-on-the-same-data-don-t-learn-the-same-features-2dsq5rc2m og:image: https://api.daily.dev/og/posts/2dsQ5Rc2M.png og:image:alt: SAEs trained on the same data don’t learn the same features og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # SAEs trained on the same data don’t learn the same features **[EleutherAI](https://daily.dev/sources/eleutherai)** · 9 min read · 0 upvotes · 0 comments ## Summary Research from EleutherAI shows that TopK Sparse Autoencoders (SAEs) trained on identical data with the same batch order but different random seeds share only ~53% of their learned latent features. Using the Hungarian algorithm to match latents via cosine similarity, the study finds that larger SAEs exhibit even less feature overlap, with 131k-latent SAEs having over 30% unshared features. Notably, many unshared latents are still interpretable, suggesting SAEs don't converge on a universal feature set. The authors attribute this to the non-convexity of the SAE loss function and connect it to feature splitting and absorption phenomena. The conclusion is that feature discovery should be viewed as a compositional, hierarchical problem rather than a search for atomic universal features. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Similar posts on daily.dev - [Neuro-Symbolic Systems as Compression, Coordination, and Alignment](https://daily.dev/posts/neuro-symbolic-systems-as-compression-coordination-and-alignment-rnad4i0y2) · Towards Data Science · 1 upvotes · 0 comments - [A Gentle Introduction to Autoencoders & Latent Space](https://daily.dev/posts/a-gentle-introduction-to-autoencoders-latent-space-x40sfxv1a) · Towards Data Science · 6 upvotes · 0 comments - [\[2512.05117\] The Universal Weight Subspace Hypothesis](https://daily.dev/posts/2512-05117-the-universal-weight-subspace-hypothesis-jrnzn939e) · Hacker News · 2 upvotes · 0 comments --- Tags: [#neural-networks](https://daily.dev/tags/neural-networks) [View this post on daily.dev](https://daily.dev/posts/saes-trained-on-the-same-data-don-t-learn-the-same-features-2dsq5rc2m) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"SAEs trained on the same data don’t learn the same features","url":"https://daily.dev/posts/saes-trained-on-the-same-data-don-t-learn-the-same-features-2dsq5rc2m","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/saes-trained-on-the-same-data-don-t-learn-the-same-features-2dsq5rc2m"},"datePublished":"2026-08-02T07:04:30.270Z","dateModified":"2026-08-02T07:32:03.018Z","description":"Research from EleutherAI shows that TopK Sparse Autoencoders (SAEs) trained on identical data with the same batch order but different random seeds share only...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/7fa3fbd6e99da7e05ab42115e7e68fde?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/7fa3fbd6e99da7e05ab42115e7e68fde?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"EleutherAI","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"EleutherAI","logo":"https://media.daily.dev/image/upload/s--aAcRyWGF--/f_auto,q_auto/v1785654263/logos/eleutherai","url":"https://daily.dev/sources/eleutherai"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/saes-trained-on-the-same-data-don-t-learn-the-same-features-2dsq5rc2m","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"neural-networks","timeRequired":"PT9M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"EleutherAI","item":"https://daily.dev/sources/eleutherai"},{"@type":"ListItem","position":3,"name":"SAEs trained on the same data don’t learn the same features"}]} ```