There is a useful engineering consequence here beyond terminology.
If intermediate tokens are not a faithful representation of the computation, then they are a pretty bad audit artifact too. We probably shouldn't be trying to make the model's internal narration more interpretable., but rather the computation around it more reproducible.
Record the actual inputs, model/version/configuration, tool observations and outputs, then make the execution replayable enough that differences between runs can be isolated.
In other words, don't ask the model to explain what it thought, and instead make the system able to show what actually happened.
Strong dislike for papers that tell me what to do in the title, especially when even the paper admits a loose correlation of the intermediate tokens compared to solution correctness.
My solutions work and they speak for themselves.
> While a human may say “aha” to indicate exactly a sudden internal state change, this interpretation is unwarranted for models which do not have any such internal state, and which on the next forward pass will only differ from the pre-aha pass by the inclusion of that single token in their context. Interpreting the “aha” moment as meaningful exemplifies the long-neglected assumption about long CoT models – the false idea that derivational traces are semantically meaningful, either in resemblance to algorithm traces or to human reasoning.
This paper addresses something that has always bothered me about LLMs. You read their reasoning, see something like “Wait, that’s wrong” and then watch them make the exact mistake they just identified.
By itself, "aha" carries no insight, but the insight is probably stated immediately after it. In that case the aha is semantically useful, by identifying the insight it is near.
it's a rhetorical heuristic that a writer should know to use when directing a reader to a declarative that they want them to pay attention to, usually because it's a non-obvious or roundabout insight
when utilized by AI, it's a probabilistic output and it's variable whether or not that rhetorical trick is useful. it also pushes a non-skeptical reader to focus too much on the following text or even to believe that they, themselves, derived some insight. this is effectively a kind of persuasive sophistry which is not helpful - adding rules around it prevents people from deluding themselves with AI
Did not read the paper so apologies if this is covered but isn't it possible that there is some recognizable semantic pattern in the training data where an "aha" is often followed by a subtle semantic shift that proves closer to the original premise in some critical way, and by emitting the "aha" token the model causes itself to produce such a subtle semantic shift that pushes the subsequent reasoning closer to the desired response?
It amounts to noise overall, but it has further unwanted and potentially misleading 'properties'. I think it's rather sobering to see how much bandwidth is still being wasted.
> but the insight is probably stated immediately after it.
If the intermediate tokens represent reasoning or thought, you would expect "aha" to occur after the thoughts that led to the realisation, including the thoughts encoding the explanation: they don't have any other state. There is no reason to draw the conclusion you've drawn. Furthermore, what LLMs are doing isn't thought.
The anthropomorphization of LLMs should be discouraged as much as possible. It perpetuates bad practices and encourages the use of these bots for tasks they are not intended for (particularly as chatbots).
Thinking traces should be treated as black boxes. There is no point in reading them. Only the LLMs’ conclusions are relevant. This is particularly true of Opus 5, which employs reasoning that seems highly questionable but very often reaches excellent conclusions (compared to its peers)
Sometimes I monitor thinking traces for misunderstandings (missing context / bad assumptions). If it's going to go off on a ~20 min task and I can catch it's going in the wrong direction in the first minute I save a lot of tokens and wasted time. I don't monitor the whole thing, mostly just the first bit to see if there was a gap or misalignment in intention.
As an aside, anthropomorphization has nothing to do with my motivations.
I think you're ending that train of thought too early. Why does this occur?
Well... We can hypothesize that these things are largely trained on internet dialogue so there's probably some correlation between threads where people are not flaming each other and the quality of the replies. They're just statistical engines so anything you can do to raise the odds of a helpful next token...
I'm essentially just making shit up here, maybe it's right, maybe it isn't, but rather than saying "it's human and we should treat it so" we're trying to get to the ground truth of how it works.
That's not a consequence of an LLM. It's a consequence of the training data. In fact, I would argue that the latest models aren't nearly as sensitive to the tone of input anymore. It's an issue that has been addressed by better curating training data.
That's my personal theory too. The model is stuffing its own context with vaguely related tokens, which helps the attention heads retrieve the right tokens.
I don't think there's anything like that going on. They just word vomit into a secondary area, and then there is an internal prompt that says "clean this up and summarize for the user".
I've been calling them film noir internal monologues, within the documents being generated by the LLM which happen to look like movie scripts.
In other words, it isn't qualitatively different from character dialogue. "Keep cheese on your pizza by using glue" is the same problem regardless of whether the script calls for the character to speak it out-loud or not.
If intermediate tokens are not a faithful representation of the computation, then they are a pretty bad audit artifact too. We probably shouldn't be trying to make the model's internal narration more interpretable., but rather the computation around it more reproducible.
Record the actual inputs, model/version/configuration, tool observations and outputs, then make the execution replayable enough that differences between runs can be isolated.
In other words, don't ask the model to explain what it thought, and instead make the system able to show what actually happened.
This paper addresses something that has always bothered me about LLMs. You read their reasoning, see something like “Wait, that’s wrong” and then watch them make the exact mistake they just identified.
when utilized by AI, it's a probabilistic output and it's variable whether or not that rhetorical trick is useful. it also pushes a non-skeptical reader to focus too much on the following text or even to believe that they, themselves, derived some insight. this is effectively a kind of persuasive sophistry which is not helpful - adding rules around it prevents people from deluding themselves with AI
If the intermediate tokens represent reasoning or thought, you would expect "aha" to occur after the thoughts that led to the realisation, including the thoughts encoding the explanation: they don't have any other state. There is no reason to draw the conclusion you've drawn. Furthermore, what LLMs are doing isn't thought.
Thinking traces should be treated as black boxes. There is no point in reading them. Only the LLMs’ conclusions are relevant. This is particularly true of Opus 5, which employs reasoning that seems highly questionable but very often reaches excellent conclusions (compared to its peers)
As an aside, anthropomorphization has nothing to do with my motivations.
And yet, they have extensive human-like behavior. If you treat them nicely or encourage them, they perform better.
Ignoring that human-like behavior is wrong headed.
Well... We can hypothesize that these things are largely trained on internet dialogue so there's probably some correlation between threads where people are not flaming each other and the quality of the replies. They're just statistical engines so anything you can do to raise the odds of a helpful next token...
I'm essentially just making shit up here, maybe it's right, maybe it isn't, but rather than saying "it's human and we should treat it so" we're trying to get to the ground truth of how it works.
I asked it for the weather. "I don't know that. I'm just a programmer."
I added "believe in yourself, you can do anything" to sysprompt, suddenly it had the confidence to Google the weather...
They are trained by gradient descent, but inference doesnt involve it.
In other words, it isn't qualitatively different from character dialogue. "Keep cheese on your pizza by using glue" is the same problem regardless of whether the script calls for the character to speak it out-loud or not.
Poster side dialogue and Q&A about this work at ICML.
https://news.ycombinator.com/item?id=49277303