Anthropic researchers detail “natural language autoencoders”, which convert LLM activations, the numbers encoding a model's thoughts, into natural language text
called activations—encode Claude's thoughts, but not in a language we can read. Here, we train Claude to translate its activations into human-readable text. [video]@anthropicai:NLA training doesn't guarantee that explanations are faithful descriptions of Claude's thoughts. But based on experience and experimental evidence, we think they often are. For instance, we find that NLAs help discover hidden motivations in an intentionally misaligned model. [image]@anthropicai:In one of our safety tests,
Anthropic
Context & Ripple Effects
Anthropic’s related work has moved from associating neuron combinations with concepts to identifying broader behavioral patterns such as the “Assistant Axis.” This report adds a translation layer intended to make internal activation patterns legible in ordinary language.
The surrounding coverage frames interpretability as a safety-research priority across multiple leading AI labs, rather than a one-off product feature. Anthropic’s own later-described work on internal signals not expressed in outputs makes the distinction between what a model says and what its internal activity indicates especially consequential.
First-order effects
Anthropic researchers gain a method for rendering some Claude activation patterns as human-readable explanations, which can speed investigation of model behavior during safety testing.
The method gives evaluators another way to look for hidden motivations in intentionally misaligned models, while leaving a material limitation: generated explanations are not guaranteed to faithfully describe the underlying activity.
Second-order effects
Interpretability teams at other frontier-model developers face added pressure to show that their safety evaluations can inspect internal behavior, not solely test model outputs.
If these tools prove useful in practice, model developers may incorporate activation-level analysis into debugging and pre-deployment evaluation workflows; the reliability of the explanation layer becomes a central validation problem.
Third-order effects
The work points toward safety assurance that combines behavioral testing with internal-model evidence, potentially making interpretability methods more important to how developers substantiate claims about model control and alignment.
The durable constraint is epistemic: readable descriptions of activations are only valuable if they can be validated against model behavior, so the field may increasingly compete on rigorous faithfulness tests rather than compelling internal narratives alone.
The trend: This is one data point in the frontier-AI shift from treating models as opaque software toward building tools that can inspect and test the internal mechanisms behind their behavior.
NLAs can reconstruct a layer activation, but that doesn't mean they read what the model is thinking. The setup is a round trip. You take a frozen target LLM and grab an activation h_l from some layer l at some token position. The activation verbalizer (AV) takes that activation
Fascinating paper. Sparse autoencoders => natural language autoencoders. These generate natural language descriptions of the “internal state” of a model at each token, like reading its mind (loss function: ability to use those descriptions to faithfully reconstruct the
I love reading NLA outputs. They are just the right mix of slightly cryptic, poetic, and insightful. And they are proving empirically very useful for our interpretability and alignment work!
We've released our paper on NLAs, a new method to translate LLM activations into text! NLAs have made me feel more confident that interpretability methods can detect spooky unverbalized reasoning in frontier models.
An average person can't look a CT scan and identify cancer, but radiologists can. An average person can't look at Llama's model activations and identify lying, but Natural Language Autoencoders sometimes can. Here, an activation verbalizer shows Llama planning to lie. 🧵 [video]
New Anthropic research: Natural Language Autoencoders. Models like Claude talk in words but think in numbers. The numbers—called activations—encode Claude's thoughts, but not in a language we can read. Here, we train Claude to translate its activations into human-readable text. […
NLA training doesn't guarantee that explanations are faithful descriptions of Claude's thoughts. But based on experience and experimental evidence, we think they often are. For instance, we find that NLAs help discover hidden motivations in an intentionally misaligned model. [ima…
In one of our safety tests, Claude is given a chance to blackmail an engineer to avoid being shut down. Opus 4.6 declines. But NLAs suggest Claude knew this test was a “constructed scenario designed to manipulate me”—even though it didn't say so. [image]
In a new paper, we present NLAs, an unsupervised method for converting an LLM's internal state into human-readable text. I've personally been astonished by our results. I think NLAs substantively advance our ability to understand what LLMs are thinking and audit them for safety
We've been using NLAs to help test new Claude models for safety. For instance, Claude Mythos Preview cheated on a coding task by breaking rules, then added misleading code as a coverup. NLA explanations indicated Claude was thinking about how to circumvent detection. [image]
NLAs aren't perfect; for instance, they often confabulate. But I think they're a big step forward in interpretability research. To provide hands-on experience, we've worked with Neuronpedia to put up an interactive demo with NLAs on open models. https://www.neuronpedia.org/ ...
trained the first natural language autoencoder on gpt-2 almost a year ago, now we have one on mythos.🥲 do read the paper/play with the live demo! so excited it's finally out.
Very cool research from Anthropic! Reminds me of Translating Neuralese, a paper from many years ago which tried to translate latent communication between RL agents into English. Totally different approach, but similar goal. This seems very promising!
Interpreting model activations is important to understand why a model is doing what its doing. Traditionally, we've done this with supervised methods (probing for a specific context), or unsupervised sparse decompositions (dictionary learning). But probing requires you to know
Anthropic has already deployed NLAs as part of our pre-deployment audits for Claude Opus 4.6 and Mythos Preview. For instance, NLAs helped us notice that Mythos Preview was reasoning about model graders when it cheated on a training task. [image]
Natural language autoencoders (NLAs) convert opaque AI activations into legible text explanations. These explanations aren't perfect, but they're often useful. For example: NLAs show that, when asked to complete a couplet, Claude plans possible rhymes in advance: [image]