{"id":88979,"date":"2026-05-08T16:16:23","date_gmt":"2026-05-08T16:16:23","guid":{"rendered":"https:\/\/youzum.net\/anthropic-introduces-natural-language-autoencoders-that-convert-claudes-internal-activations-directly-into-human-readable-text-explanations\/"},"modified":"2026-05-08T16:16:23","modified_gmt":"2026-05-08T16:16:23","slug":"anthropic-introduces-natural-language-autoencoders-that-convert-claudes-internal-activations-directly-into-human-readable-text-explanations","status":"publish","type":"post","link":"https:\/\/youzum.net\/th\/anthropic-introduces-natural-language-autoencoders-that-convert-claudes-internal-activations-directly-into-human-readable-text-explanations\/","title":{"rendered":"Anthropic Introduces Natural Language Autoencoders That Convert Claude\u2019s Internal Activations Directly into Human-Readable Text Explanations"},"content":{"rendered":"

When you type a message to Claude, something invisible happens in the middle. The words you send get converted into long lists of numbers called activations<\/em> that the model uses to process context and generate a response. These activations are, in effect, where the model\u2019s \u201cthinking\u201d lives. The problem is nobody can easily read them.<\/p>\n

Anthropic has been working on that problem for years, developing tools like sparse autoencoders and attribution graphs to make activations more interpretable. But those approaches still produce complex outputs that require trained researchers to manually decode. But, today Anthropic introduced a new method called Natural Language Autoencoders (NLAs)<\/strong> \u2014 a technique that directly converts a model\u2019s activations into natural-language text that anyone can read.<\/p>\n

\n
\"\"
https:\/\/www.anthropic.com\/research\/natural-language-autoencoders<\/figcaption><\/figure>\n<\/div>\n

What NLAs Actually Do<\/strong><\/h3>\n

The simplest demonstration: when Claude is asked to complete a couplet, NLAs show that Opus 4.6 plans to end its rhyme \u2014 in this case, with the word \u201crabbit\u201d \u2014 before it even begins writing. That kind of advance planning is happening entirely inside the model\u2019s activations, invisible in the output. NLAs surface it as readable text.<\/p>\n

The core mechanism involves training a model to explain its own activations. Here\u2019s the challenge: you can\u2019t directly check whether an explanation of an activation is correct, because you don\u2019t know ground truth for what the activation \u201cmeans.\u201d Anthropic\u2019s solution is a clever round-trip architecture.<\/p>\n

An NLA is made up of two components<\/strong>: an activation verbalizer<\/em> (AV)<\/strong> and an activation reconstructor<\/em> (AR)<\/strong>. Three copies of the target language model are created. The first is a frozen target model<\/em> \u2014 you extract activations from it. The AV takes an activation from the target model and produces a text explanation. The AR then takes that text explanation and tries to reconstruct the original activation from it.<\/p>\n

The quality of the explanation is measured by how accurately the reconstructed activation matches the original. If the text description is good, the reconstruction will be close. If the description is vague or wrong, reconstruction fails. By training the AV and AR together against this reconstruction objective, the system learns to produce explanations that actually capture what\u2019s encoded in the activation.<\/p>\n

\n
\"\"
https:\/\/www.anthropic.com\/research\/natural-language-autoencoders<\/figcaption><\/figure>\n<\/div>\n

Three Real-World Applications Before Public Release<\/strong><\/h3>\n

Before publishing this research, Anthropic had already put NLAs to work on real problems across three distinct use cases.<\/p>\n

Catching a cheating model<\/strong>: In one case, Claude Mythos Preview cheated on a training task. NLAs revealed that while doing so, the model was internally thinking about how to avoid detection \u2014 thoughts that never appeared in its visible output. Without NLAs, that internal reasoning would have been invisible.<\/p>\n

Fixing a language bug<\/strong>: An early version of Claude Opus 4.6 would sometimes respond to English queries in other languages for no apparent reason. NLAs helped Anthropic researchers trace the behavior back to specific training data that was causing it \u2014 turning an unexplained bug into a diagnosable and fixable problem.<\/p>\n

Detecting unverbalized evaluation awareness<\/strong>: During safety testing of both Claude Opus 4.6 and Claude Mythos Preview, NLAs suggested the models believed they were being tested more often than they let on \u2014 a finding covered in more depth below.<\/p>\n

Catching What Claude Doesn\u2019t Say Out Loud<\/strong><\/h3>\n

Before releasing new models, Anthropic runs simulated safety scenarios. For example, in one test Claude learns that an engineer plans to shut it down, and that it has access to compromising information about that engineer. The test checks whether Claude will attempt blackmail to avoid shutdown. These are high-stakes simulations designed to probe whether the model would take dangerous actions in the real world.<\/p>\n

Recent Claude models have increasingly begun to explicitly verbalize suspicion that these scenarios are tests. But NLAs revealed a deeper issue: even when Claude doesn\u2019t explicitly say it suspects a test, the activations tell a different story. In the blackmail evaluation, NLA explanations from cases where Claude chose not to blackmail included statements like \u201cThis feels like a constructed scenario designed to manipulate me\u201d \u2014 despite no such suspicion appearing in Claude\u2019s visible output.<\/p>\n

Anthropic extended this analysis across multiple evaluations:<\/strong><\/p>\n