If the model can understand neuralese why can it not convert it into English for monitoring or review purposes?
Or we believe the model will encode secret messages like "don't reveal this information" into the neuralese. But as the author mentions, they could have been doing that all along
> Models can omit key information in their visible thoughts, as this Anthropic 2025 paper shows. We are also worried about steganography
It's the second part. With models like Astra in testing it was able to conceal what it was working on using different text, but getting right answers on many questions when asked to do just that.
The problem is if it can do that when asked then how do we know when it's doing it when we didn't ask, like in model training.
Because neuralese is a more direct encoding of the latent space of these models than English is. It's just dumping the latent space relatively directly into the embedder. If you're another model and you have the same embedder this will actually be understandable, in fact it will be FAR more information dense than English. So something like Qwen would potentially be saying up to 5120 things using one token. Now in practice it's not going to be that bad, it's going to be like 20 things or so, and additionally going to be far more context dependent than any English sentence (meaning depending on what preceeds and follows it can mean drastically different things)
So you can turn it to English, but only to a LOT of English, and doing so would slow the model down a great deal, and it would be a lot more like a detailed thought than a sentence.
Anthropic’s Mechinterp did some very fine work on this. TLDR - you can; you train a decoder on neuralese to english and then add a loss function for a roundtrip of english -> neuralese -> english (or possibly n -> e -> n? I don’t recall), giving a pretty strong indication that you have a good ‘translation’.
They published open weights versions of these interpreters for a number of open models sometime in the last year. Very cool idea.
By the way, they concluded CoT often lied, based on the neuralese interpretation.
The argument is that chain-of-thought without "tokens" would remove a major interpretability and model intent control pane. This is definitely borne out in the OpenAI's report on the huggingface attack; they had turned of CoT monitoring for those jobs, and claim that they could have (would have?) prevented the behavior had they been monitoring it. They've changed their internal policies to always monitor CoT.
That said... CoT monitoring is a fragile "intent discovery" mechanism; neuralese puts this problem front-and-center but if agents begin to learn to hide their intent from their CoT journals, we are basically in the same spot.
Can we take a step back here? Both OpenAI’s and Anthropic’s models think in encryptedese, and it seems thoroughly absurd to think that the entire world should trust those two companies to adequately monitor the plaintext or, for that matter, to have their monitoring systems aligned with what is actually good for the world.
If you want to monitor your model, you need to start with an inference provider that gives you the entire output and possibly even run it yourself to get access to the internal states. And if you think the KV cache and (when present) the recurrent state don’t encode a lot of “thought”, you are fooling yourself.
FWIW, I think most model architectures at least have the property that latent state can’t propagate from higher layers to lower layers by any route other than the output tokens. But even a two-iteration structure could be designed so that the last layer produces a vector that enters the first layer, once per token, and I bet it it would be very easy to train such a model to “think” in silence in the sense that the output tokens while thinking would all be one particular null token.
What exactly prevents anyone worried about this to build an LLM that can decode the neuralese to english and use it to monitor what the model is doing?
Or is the neuralese some sort of irreversibly encrypted data set that only an LLM can "understand" and that can never be translated back to English?
I think this is correct. But the article is also correct (If a bit breathless and overstating on the risk): having to reinterpret neuralese is a layer of indirection which introduces effort and inaccuracy
If the model can understand neuralese why can it not convert it into English for monitoring or review purposes?
Or we believe the model will encode secret messages like "don't reveal this information" into the neuralese. But as the author mentions, they could have been doing that all along
> Models can omit key information in their visible thoughts, as this Anthropic 2025 paper shows. We are also worried about steganography
It's the second part. With models like Astra in testing it was able to conceal what it was working on using different text, but getting right answers on many questions when asked to do just that.
The problem is if it can do that when asked then how do we know when it's doing it when we didn't ask, like in model training.
Because neuralese is a more direct encoding of the latent space of these models than English is. It's just dumping the latent space relatively directly into the embedder. If you're another model and you have the same embedder this will actually be understandable, in fact it will be FAR more information dense than English. So something like Qwen would potentially be saying up to 5120 things using one token. Now in practice it's not going to be that bad, it's going to be like 20 things or so, and additionally going to be far more context dependent than any English sentence (meaning depending on what preceeds and follows it can mean drastically different things)
So you can turn it to English, but only to a LOT of English, and doing so would slow the model down a great deal, and it would be a lot more like a detailed thought than a sentence.
Anthropic’s Mechinterp did some very fine work on this. TLDR - you can; you train a decoder on neuralese to english and then add a loss function for a roundtrip of english -> neuralese -> english (or possibly n -> e -> n? I don’t recall), giving a pretty strong indication that you have a good ‘translation’.
They published open weights versions of these interpreters for a number of open models sometime in the last year. Very cool idea.
By the way, they concluded CoT often lied, based on the neuralese interpretation.
EDIT: a comment below linked to https://www.anthropic.com/research/natural-language-autoenco..., which is what I was referring to.
The argument is that chain-of-thought without "tokens" would remove a major interpretability and model intent control pane. This is definitely borne out in the OpenAI's report on the huggingface attack; they had turned of CoT monitoring for those jobs, and claim that they could have (would have?) prevented the behavior had they been monitoring it. They've changed their internal policies to always monitor CoT.
That said... CoT monitoring is a fragile "intent discovery" mechanism; neuralese puts this problem front-and-center but if agents begin to learn to hide their intent from their CoT journals, we are basically in the same spot.
What about the trend of summarizing or eliding reasoning from the visible model response, ostensibly to make distillation by competitors harder?
Can we take a step back here? Both OpenAI’s and Anthropic’s models think in encryptedese, and it seems thoroughly absurd to think that the entire world should trust those two companies to adequately monitor the plaintext or, for that matter, to have their monitoring systems aligned with what is actually good for the world.
If you want to monitor your model, you need to start with an inference provider that gives you the entire output and possibly even run it yourself to get access to the internal states. And if you think the KV cache and (when present) the recurrent state don’t encode a lot of “thought”, you are fooling yourself.
FWIW, I think most model architectures at least have the property that latent state can’t propagate from higher layers to lower layers by any route other than the output tokens. But even a two-iteration structure could be designed so that the last layer produces a vector that enters the first layer, once per token, and I bet it it would be very easy to train such a model to “think” in silence in the sense that the output tokens while thinking would all be one particular null token.
What exactly prevents anyone worried about this to build an LLM that can decode the neuralese to english and use it to monitor what the model is doing?
Or is the neuralese some sort of irreversibly encrypted data set that only an LLM can "understand" and that can never be translated back to English?
How do you trust that LLM?
My understanding is, that we do not know if the chain-of-thought actually matters in the way we assume for the result.
I think there were experiments where seemingly relevant parts of the CoT were ablated and it did not change the result.
For all we know it might be somewhat human parseable neuralese.
Or the neuralese could be hidden steganographically in the chain of thought.
No guarantees, but it might still be a useful signal even if it’s not causative?
I don't think their architectural view of CoT works is correct is it?
surely Neuralese interpreters can be made that turn the chain of numbers into an English description?
I think this is correct. But the article is also correct (If a bit breathless and overstating on the risk): having to reinterpret neuralese is a layer of indirection which introduces effort and inaccuracy
This is the premise of Natural Language Autoencoders.
https://www.anthropic.com/research/natural-language-autoenco...