Steganography in Chain of Thought Reasoning
5Hoagy
3A Ray
1Hoagy
1Vaniver
3A Ray
1Quintin Pope
2A Ray
0Quintin Pope
2A Ray
2Quintin Pope
1A Ray
New Comment
For the ELK contest I was working on something similar, but where the problem is more acute (you're trying to force a MuZero-style reinforcement learner to compress its hidden state into words).
The methods I proposed to deal with this are:
- Cut down the null space by training a language model to rewrite the written words in a different format which preserves the meaning as closely as possible, so that the CoT reasoner is forced to respond to similar sentences in similar ways.
- Use multiple models which are trained separately but have to respond to each others' thoughts, making it more difficult to have a shared steganographic protocol.
- Use something like dropout on the chains of thought, so that it can't hide information in the combination of multiple sentences.
- Have some non-linguistic bits which we don't apply optimization to, but instead focus our intepretability efforts on. The hope is that this soaks up the highest-order bits that aren't easily expressible in language,so there's less need to hide them within words. Applying dropout to these is particularly necessary.
1Putting what I think is the most important part of my reply first: I think research into mitigations is premature and instead demonstrating/measuring the phenomena should take priority.
However given that, I think I agree that these are all possible mitigations to the phenomena, in particular (rephrasing your points):
- Reducing the channel capacity, and thus the amount of information that can be transmitted
- Constraining the outputs, and thus reducing the degrees of freedom of the null space
- Adding noise (via dropout or otherwise) also reduces the channel capacity, and probably also has some desirable side effects (like increasing the diversity or creativity of reasoning)
- Introducing "barriers" past which optimization for hidden information becomes difficult (in your example, you use different models; I also think "stop gradients" in end-to-end training has similar effects, if weaker)
- If you can detect/localize this phenomena, you might be able to use interpretability like techniques to limit or constrain it
- I'm less certain about this, but possible future interpretability research will unlock this ability.
Also I think an important point here is that we can expect to see it with fairly simple techniques -- merely including sampled/generated data in the training set is sufficient, as opposed to it requiring a complicated reinforcement learning algorithm like MuZero.
Did you publish your proposal? I'd be interested in reading it.
Yeah I agree demonstrating it and examining it is very important, especially in chain-of-thought where the existence of hidden information isn't a certainty.
They're a bit out of date but the original proposal is ELK Sub - Note-taking in internal rollouts and the multiple model protocol is in Note Taking Without Hidden Messages.
Haven't done any proper testing but it's high up on my agenda, would be interested in working out what tests would be best - will elaborate later.
For fixed-duration reasoning chains, or in situations where the AI is trying to facilitate cooperation between different passes (like in the proposed scenario where it gets information but then loses access to it), this seems important. For example, in a situation where you show the model a passage, let it make notes, and then have the same model look just at the notes and answer a reading comprehension question, I expect that likely it will want to embed information in the null space of the notes. [Or in an HCH-style situation, it might want to embed the original question into subquestions, and might do so steganographically, in a way that breaks some of the things one might be hoping for there.]
But I'm somewhat confused about whether you should expect it to crop up in recursive prompting cases (probably you don't?). If you secretly know the answer in the first pass, don't you want to just say the answer and terminate? If you secretly know a relevant fact in the first pass, don't you want to openly express it and recur? I don't see how steganography is rewarded over openness. [I'm setting aside the case where it invents a new language to express non-linguistic concepts more cheaply, and assuming we're in the natural language setting where standard language is the 'natural' way to transmit the information.]
I think this is worth considering even in cases of things like recursive prompting for a couple of reasons that I haven't organized:
- In general additional compute time lets you improve your answer, so by default I'd expect answers to any significant question to take multiple steps
- I agree for trivial questions it might not do any of this. My expectation is that AI deployments impacts will mostly be on things like non-trivial questions or tasks
- So far current research hasn't done a great deal of regularizing or penalizing models for using a lot of thinking
- Mostly they've optimized for "maximize the chance of correct answer"
- This could change in the future, especially in cases where the model is expected to be deployed / otherwise sensitive to runtime costs
- Steganography is rewarded over openness if the openness is somehow far from the prior for a given piece of information it wants to transmit
- Caveat this point is more speculative than the above document
- One of my points is that due to null space + optimization pressure, it's very easy and convenient for gradient descent to put information in these hidden dimensions
- I think one of the things this can do well is handle concepts that don't yet exist -- and so reasoning moves like "invent a new term" or "give an illustrative example" or "invent a thought experiment" are lower likelihood under the prior
- In these cases, it can be more likely for steganography to transmit certain information, if it is less expensive in terms of KL penalties
- I do agree that for straightforward reasoning on straightforward evidence, where the reasoning is human-like and close to the prior, we can expect it to mostly reason out loud in a way that is "open"
- Myopia -- though less important than it was earlier in RFLO research -- seemed like a good alignment criterion to aim for, and still does today somewhat
- The original starting point for this direction of thinking was about Myopia / Non-Myopia in chain of thought reasoning systems
- The sorts of guarantees I was hoping to get for chain of thought systems included myopia guarantees
- Steganography came up as an example of how it could break myopia (other more complicated examples include side effects and situational awareness)
- The concepts in this document came up by asking "how might steganography arise naturally given current research techniques and directions"
I think this is a very important direction to investigate. CoT reasoners seem like our best shot at aligning an AGI.
The KL objective pushes these correlations into the “null space” where they are not penalized.
I don’t think this is true. There are many ways to express the same underlying idea, but GPT-3 is not agnostic between them. Divergences from GPT-3’s nullspace are still penalized. KL isn’t wrt “does this text contain the same high level contents according to a human?” It’s wrt “does this text match GPT-3’s most likely continuations?”
Edit: not actually how KL divergence works.
I think I understand what you're saying, but I want to double check and try laying it out explicitly.
- The "null space" is actually the null space of the prior model, which is human-like, but not exactly lined up to human perception of language
- Because of this, the information is "hidden" w.r.t. human perceptions, it's "hidden" w.r.t. the language model prior
I think I agree with both of these points, but here's my thinking for why I still expect to see this phenomena (and why the article above was simplified to just be "human")
- Empirically drawing many samples results in lots of "contains similar high level contents according to a human (me)"
- It might be worth studying this explicitly, such as setting up a bunch of chain of thought reasoning like scenarios, and sampling a lot of generated reasoning, then asking humans to rate how similar the generations are
- My prediction is that people will think that they are largely similar
- We could also use this same method to study human-measured relevance of other non-trivial sampling methods, like beam search, etc
- "Hiding" w.r.t. language model prior is sufficient, in the limit of good language models
- One of the things I think my reasoning relies on is that we will continue to produce better language models, which means the prior over time more and more closely resembles human priors
- This suggests that one way to avoid this phenomena is to only use smaller / less capable models -- or at least they will do a worse job at hiding information. I currently expect that many research labs will continue to do chain of thought reasoning with state of the art models, though.
- There's a lot of room in the model distribution
- Chain of thought reasoning is not currently, nor do I expect it will ever be, overconstrained.
- Instead, it can be seen as a mechanism for dramatically increasing the space of possible outputs, both those seen during training, and those encountered at test time
- Human preference effects
- Research Campaign Effects: Human researchers will want human-like outputs, and thus will be expected to tweak/retrain/etc in order to get human-like reasoning in the chain of thought. This introduces a separate optimization pressure from the KL penalty towards reasoning being human-like
- Human preference models: I think an easy-to-predict future direction here is training reward models to match subjective human preferences and then optimizing with them or sampling with them (e.g. best of N) -- these preference models could also introduce additional optimization pressure towards human-like
I think the most important point in all of this is that I expect the model prior to be sufficiently close to the human prior -- and I'd be interested in learning more about evidence or research to the contrary.
Apologies for my first comment. I was tired, and most of what I said about KL divergence is actually wrong.
The core issue is that you can't actually avoid the KL penalty by hiding in the nullspace. That's just not how KL divergence works. A KL divergence of zero implies that the two LMs specify exactly the same distribution over tokens. If you deviate away from the reference LM's prior in any manner, then you take a KL penalty.
E.g., if the reference LM is indifferent between the phrases “My stomach is grumbling” or “I could use a snack”, but the steganographic LM always uses “My stomach is grumbling”, then this deviation is also penalized by KL divergence, regardless of whether the reference LM's indifference between the two phrasings means you call this part of the LM's "nullspace".
If what you’re saying is “any change to the distribution will change KL” — I think that’s just correct.
This also applies to changes during training where the model is learning to perform better on the objective task.
So we are expecting some amount of KL divergence already.
My claims are:
-
The cheapest place to hide information (due to KL) are places where the model already has high entropy (ie it is uncertain between many possible outputs)
-
optimization pressure will try to push this extra information into the cheapest places to hide
-
the increase in KL won’t be clearly distinguishable from the increase due to increased performance on the task
I think it probably works out in the other direction. The easiest way to encode problem-relevant info is by changing the conceptual content of the generated text, not weird changes to the grammatical rules. This is because KL penalties and cross entropy loss actually prioritize modeling exact grammatical patterns over conceptual content. E.g., GPTs will quickly learn to generate ~perfect grammar, but are slow to pick up on / produce coherent conceptual content.
Another framing is to think of the KL regularizated capabilities training as Bayes updating of the model’s priors on the distribution over text. The model should have much stronger priors on grammatical rules than on higher level concepts. So, the update probably changes conceptual patterns more than grammatical ones.
the increase in KL won’t be clearly distinguishable from the increase due to increased performance on the task
Can’t you just compare the token predictions of the new model against the reference LM? Then you can check if they disagree on weird grammatical patterns or on conceptual tokens that are related to a human-like understanding of how to tackle the problem. Maybe you can even bias the KL penalty to apply higher weight to stopwords, punctuation, and synonym choices?
I think at this point these feel like empirical questions, which I think would be much more clearly answered by demonstrations or experiments.
Trying to encode an additional penalty on changing non-semantic information is an interesting idea.
However I think you're missing that you don't have the ability to directly compare to a reference LM in cases where you're training to improve on some performance benchmark. During training the model will change its predictions on everything to some degree (both semantic and nonsemantic content).
So your proposed detection will always show some amount of disagreement between the prior and the trained model on weird grammatical patterns as well as conceptual tokens. The question is: "is the difference merely due to the changes to improve performance, or is it also transmitting hidden information"