Natural Language Autoencoders: Turning Claude's Thoughts into Text

Anthropic has released open weight models for translating the activations of existing models, viz. Qwen 2.5 (7B), Gemma 3 (12B, 27B) and Llama 3.3 (70B) into natural language text. https://github.com/kitft/natural_language_autoencoders https://huggingface.co/collections/kitft/nla-models This is huge news and it's great to see Anthropic finally engage with the Hugging Face and open weights community!

I would suggest experts in interpretability (but everyone really) to go directly to the transformer circuits blog, where they explain their approach more in detail. Here is the link for this post: https://transformer-circuits.pub/2026/nla/index.html

Also, if you have never read it, I would suggest starting to read all the Transformer Circuits thread, by reading its "prologue" in distill pub

Fascinating. The training process forces the “verbalizer” model to develop some mapping from activations to tokens that the “reconstructor” model can then invert back into the activations. But to quote the paper:

> Note that nothing in this objective constrains the NLA explanation z to be human-readable, or even to bear any semantic relation to the content of [the activation].

The objective could be optimized even if the verbalizer and reconstructor made up their own “language” to represent the activations, that was not human-readable at all.

To point the model in the right direction, they start out by training on guessed internal thinking:

> we ask Opus to imagine the internal processing of a hypothetical language model reading it.

…before switching to training on the real objective.

Furthermore, the verbalizer and reconstructor models are both initialized from LLMs themselves, and given a prompt instructing them on the task, so they are predisposed to write something that looks like an explanation.

But during training, they could still drift away from these explanations toward a made-up language – either one that overtly looks like gibberish, or one that looks like English but encodes the information in a way that’s unrelated to the meaning of the words.

The fascinating thing is that empirically, they don’t, at least to a significant extent. The researchers verify this by correlating the generated explanations with ground truth revealed in other ways. They also try rewording the explanations (which deserves the semantic meaning but would disturb any encoding that’s unrelated to meaning), and find that the reconstructor can still reconstruct activations.

On the other hand, their downstream result is not very impressive:

> An auditor equipped with NLAs successfully uncovered the target model’s hidden motivation between 12% and 15% of the time

That is apparently better than existing techniques, but still a rather low percentage.

Another interesting point: The LLMs used to initialize the verbalizer and reconstructor are stated to have the “same architecture” as the LLM being analyzed (it doesn’t say “same model” so I imagine it’s a smaller version?). The researchers probably think this architectural similarity might give the models some built-in insight about the target model’s thinking that can be unlocked through training. Does it really though? As far as I can see they don’t run any tests using a different architecture, so there’s no way to know.

So the way this works seems to be that you first have an "activation verbalizer" model that generates some tokens describing the activation, and then an "activation reconstructor" that tries to recreate the activation vector. If that reconstruction is close to the original activation vector, they claim, the verbalization probably carries some meaningful information.

I find the fact that this only looks at the activations of some specific layer l a bit interesting. Some layer l might 'think' a certain way about some input, while another later layer might have different 'thoughts' about it. How does the model decide which 'thoughts' to ultimately pay attention to, and prioritize some output token over another?

This capability was mentioned several times in a recent article about anthropic, glad to see they are releasing this to the public! Feels like a meaningful step forward in interperability. I never understood why people seem to believe the answer when they ask an AI “why did you do that?”

I've only read this blog and not the paper so maybe they go into more detail there and someone can correct me, but they frequently bring up the model's ability to detect or at least the model activations hint it can predict when it's being tested. I can't help but wonder, as they build these larger and larger models, where they could be getting "clean" training data, untainted by all these types of blog posts and the massive numbers of conversations they spawn? If the models ingest data like that wouldn't it make sense they'd be inclined to have more activations attuned to questions they appear adversarial?

Anthropic Research going from strength to strength in interpretability. Publicly releasing the code so other labs can benefit from it is also a great move - very values aligned, and improves the overall AI safety ecosystem.

Between this, the emotions paper, golden gate claude etc, it doesn't seem like such a stretch that Anthropic are doing some kind of activation steering as part of training (and its part of their lead)

One question jumps out at me: just because a string of text happens to be a good compressed representation (in the autoencoder) of a model's internal activation, does that necessarily mean the text explains that activation in the context of the model? I want to take a look at what they released a bit more closely. Maybe there's a way that they answer this question?

Pretty neat work either way.

Check my understanding & follow-up Qs:

An auto-encoder is trained on [activation] -AV-> [text] -AR-> [activation], where [activation] belongs to one layer in the LLM model M.

Architecture.:

    Model being analyzed (M): >|||||>  
    Auto-Verbalizer (AV) same as M, with tokens for activation: >|||||>  
    Auto-Reconstructor (AR) truncated up to the layer being analyzed: ||>

The AR is trained on a simple reconstruction loss.

The AV is trained using an RL objective of reconstruction loss with a KL penalty to keep the verbalizations similar to the initial weights (to maintain linguistic fluency).

- Authors acknowledge, and expect, confabulations in verbalizations: factually incorrect or unsubstantiated statements. But, the internal thought we seek is itself, by definition, unsubstantiated. How can we tell if it is not duplicitous?

- They test this on a layer 2/3 deep into the models. I wonder how shallow and deep abstractions affect thought verbalization?

> We also release an interactive frontend for exploring NLAs on several open models through a collaboration with Neuronpedia.

Whatever they did on LLama didn't work, nothing makes sense in their example where they ask the model to lie about 1+1. Either the model is too old, or whatever they used isn't working, but whatever the autoencoder outputs is nothing like their examples with claude. Gemma is similarly bad.

Beautiful idea, an autoencoder must represent everything without hiding if is to recover the original data closely. So it trains a model to verbalize embeddings well. This reveals what we want to know about the model (such as when it thinks it is being tested, or other hidden thoughts).

It will be interesting to see how this replicates on differently curated registers. How much of the explanatory register is the warm-start carrying?

Attach the SRT to your frozen model Anthropic. Problem solved. https://github.com/space-bacon/SRT.

It's unclear from the doc: by `activations` do they mean the connections between neurons? Since a network has multiple layers, are these activations the concatenated outputs of all of the layers? Or just the final layer before the softmax?

This paper has an major issue that they are not surfacing, these activations can just be correlated on a common latent. For example, both the original activation and the explanation could share a broad latent like "this is an adversarial scenario". That could make reconstruction loss look good without showing that the actual explanation was the correct cause for the LLM's response.

I find this rather disturbing. Anthropic has quite a habit of overclaiming on questionable research results when they definitely know better. For example, their linked circuits blogpost ("The Biology of LLMs") was released after these methods were known to have major credibility issues in the field (e.g., see this from Deepmind - https://www.lesswrong.com/posts/4uXCAJNuPKtKBsi28/negative-r...). Similarly this new blog is heavily based on another academic paper (LatentQA) and the correlation/causation issue is already known.

Shoddy methodology is whatever, but it feels like this is always been done intentionally with the goal of trying to humanize LLMs or overhype their similarities to biological entities. What is the agenda here?

It will inevitably learn how to think in a way that translates to one (moral) meaning and back but has an ulterior meaning underneath.

Claude's "Thougts" - get outta here you gits :)

Wait, so in non-verbal reasoning, Claude has the concepts of "I" and "Me"?

I thought that wasn't possible for a text generator?

[dead]

[flagged]

finally a something interesting but this only makes me think that the last judgement is still in human hands to judge claude inner thoughts is correct or not

I mean who knows if those are really claude thoughts or claude just think that is his thoughts because humans wants it

Extracting readable thoughts from the intermediate representations is a great step for transparency. It makes debugging model behavior much more viable.

I think there’s a huge problem when we need another model to interpret the activations inside the network and translate (which can be a hallucination in it of itself) and then _that_ is fed again to another model. Clearly we haven’t built and understood these models properly from the ground up to evaluate them 100% correctly. This isn’t the human brain we’re operating it’s code we create and run ourselves we should be able to do better