Of course not. Because the article uses the words 'thought' and 'reasoning' and even 'faithful' to mean something other than their normal meanings, but then expects them to behave exactly the same.
Every field has terms of art, and 'reasoning' is one for LLMs. But that doesn't mean it has the same properties as 'reasoning' in other contexts, because you're not referring to the same thing.
Related: Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces![1]
> Our findings consistently challenge the prevailing narrative that intermediate tokens constitute a semantically meaningful reasoning process. First, we observe a pronounced lack of correlation between solution correctness and trace validity—models frequently produce invalid reasoning traces even when they arrive at correct solutions. Second, and more strikingly, models trained on corrupted or semantically irrelevant traces achieve performance comparable to, and often exceeding, that of models trained on correct traces, especially on out-of-distribution tasks.
The author of this paper is in ASU and does a lot of excellent work in this space. People should check it out. Especially the paper titled "Beyond Semantics..." His twitter is also active _and_ high SNR.
Over time, I've learned to accept that many people -- even very clever ones -- are incapable of holding a metaphor at arm's length. Once they accept the words of a metaphor as applicable at all, the metaphor collapses entirely into literalism for them. They can no longer see that the metaphor was just a tool with inherent limitatation and boundaries.
Because the field of artificial "intelligence" is constructed around the idea of applying psychological metaphors to computational systems (a very powerful idea!) it's almost a worst case scenario for these people.
Suddenly, they're reversing the metaphors and applying computational schema to psychological processes ("aren't we really just stochastic parrots ourselves?!"); or, like here, they find themselves surprised and confused when they stumble across the natural boundaries of the metaphor experimentally.
It's because they never had sight of the boundaries in the first place and maybe never can quite see them. The words only make sense to them as literal equivalence, and so their surprise when they run into stuff like this is earnest and deep.
I don't think the boundary between "generalization" and "metaphor" is very well defined. When you go from an exemplar of 1 to 2, you're going to find all kinds of edge cases where attributes of the thing being demonstrated that had seemed to be essential turn out to not be necessary.
I think you certainly could look at LLMs as "thinking" metaphorically, but I also don't think it is necessarily only a metaphor.
This paper puts words to something I’ve noticed repeatedly with LLMs, particularly Qwen3.6. When I read its reasoning, it appears to recognise the mistake and then carry on as if it hadn’t noticed it at all.
> models often determine their answers based on implicit biases tied to question templates, then construct reasoning chains to justify their predetermined conclusions
> its reasoning was correct right until the final step (Yes/No answer)
First author here, surprised to see the paper in HN! :)
When doing the paper we noticed that models are very good at generating post-hoc plausible CoT, which to me knowledge can happen quite often with relatively easy tasks.
I agree, and I very strongly dislike it, to be polite about it. It contributes absolutely nothing and is an excellent way of hand-waving away literally anything an AI model does. Saying "well people do this too" is a great way to rationalize away anything you can imagine that an AI model would be capable of, because "humans do it too so what's the big deal, guys?"
I can just immagine the response to a headline "LLM chooses mass death: thousands killed in horrific AI accident" being something like "lots of humans have caused mass death too."
It's the grounded portion of a feedback loop searching for the 'why is this happening' thinking. I'd imagine most of my own comments in this area boil down to "GIGO" most of the time.
It's a relevant comment in this instance because we're discussing concepts you need to be both trained and practiced in to reason about, and that our discipline has traditionally been blind to. Plenty of people working with LLM context issues who've never been exposed to the idea of 'subtext' or could tell you why it would matter to their direction of effort.
The current political climate is well reasoned and intentional. It might not be yours or mine, however the system is working exactly as the ones paying for it have intended.
My brother and I have been arguing about that all of our lives. He believes everything is intentional and it's just a matter of discovering who benefits. I see chaos that nobody intends or controls. His political landscape is a tapestry of conspiracy theories and mine is a fog of war. I think his is more comforting, since it admits a possibility of a rational, predictable world.
I'd synthesize those as: "There is a lot of chaos with no central plan, but every small piece happens because someone believes they will benefit."
In other words, a lot of this depends on what scale/scope is being inspected. On the high level, the world is chaos rather than a meticulous and inscrutable plan of the Illuinati Shadow Cabal. On the low level, people do things for reasons, even if they're dumb ones.
With respect to the "current political climate", I'd like to suggest that a lot of dumb or "against their own interests" stuff is due to people prioritizing costly in-group loyalty signals. Their tribal interest in staying good with the club is higher than their interest against a dumb national policy.
It’s intentional in the sense that actors are acting intentionally for their own benefit (or at least what they believe is beneficial) and following incentives. Not that there is a master planner who manipulates everything
The current admin seems to have quite a few long term plans they have been working towards.
Project 2025, Maralago accords.
So far the only major policy item the Trump admin seems to have not intended was the Iran War. Israel killing the intended replacement, Iran leveraging the straight of Hormuz, and dropping three Tomahawks on an elementary school really botched that one.
I see this when just using some chat bot that shows the “reasoning” steps (ad-hoc observation of course, it’s really cool that people are actually studying it).
It is annoying when the bot seems be “reasoning” correctly and then makes an obvious mistake at the end. And perplexing when it seems to be completely wrong and then pull the right answer out of a magic hat at the end.
I guess it makes sense; the “reasoning” steps aren’t actually doing logic, just adding more context to influence the final generation, right? But it is weird to see.
This paper made me wonder not whether the chain we can read is actual “thought,” but what conclusions we can draw by observing it. It is an output channel, but not direct access to the black box that creates it. The question pairs present an interesting experiment, but they also got me thinking about semantics: the same underlying relation can have many valid representations, and how models reason across those representations can tell us more about their stability and correctness. Basically, are models semantically consistent when given different representations of the same underlying relation? Do their answers transform as the relationship requires, and do their explanations remain consistent with that relationship?
It's... complicated. Yes, RL reward hacking makes it learn "bird language" and yes, reasoning traces can be misleading. However they also pretty clearly steer the final reply and not simply justify it, and can stay somewhat coherent and relevant with readability SFT and rewards. All these phenomenas coexist, they aren't mutually exclusive. Reasoning traces are still useful for debugging.
It is and has been - skipping the tokens causes performance to drop. But replacing the tokens with filler causes the performance to drop, but by a lot less. You can also train models to emit broken or unrelated thinking tokens and their performance is also not much worse than the ones that are trained to output somewhat coherent thinking traces.
This points to a hypothesis that the content of the tokens is only slightly related to the mechanism by which it improves performance, and that primarily the extra tokens allow the original prompt to be processed more deeply by the model, because earlier tokens will essentially pass through the model many more times than later ones.
Of course not. Because the article uses the words 'thought' and 'reasoning' and even 'faithful' to mean something other than their normal meanings, but then expects them to behave exactly the same.
Every field has terms of art, and 'reasoning' is one for LLMs. But that doesn't mean it has the same properties as 'reasoning' in other contexts, because you're not referring to the same thing.
Why doesn't my asteroid belt buckle?
Related: Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces![1]
> Our findings consistently challenge the prevailing narrative that intermediate tokens constitute a semantically meaningful reasoning process. First, we observe a pronounced lack of correlation between solution correctness and trace validity—models frequently produce invalid reasoning traces even when they arrive at correct solutions. Second, and more strikingly, models trained on corrupted or semantically irrelevant traces achieve performance comparable to, and often exceeding, that of models trained on correct traces, especially on out-of-distribution tasks.
[1]: https://arxiv.org/abs/2504.09762
Also posted to HN today with a couple comments: https://news.ycombinator.com/item?id=49360140#49363374
The author of this paper is in ASU and does a lot of excellent work in this space. People should check it out. Especially the paper titled "Beyond Semantics..." His twitter is also active _and_ high SNR.
Related: Poster side dialogue and Q&A about this work at ICML. Very good. https://news.ycombinator.com/item?id=49277303
Over time, I've learned to accept that many people -- even very clever ones -- are incapable of holding a metaphor at arm's length. Once they accept the words of a metaphor as applicable at all, the metaphor collapses entirely into literalism for them. They can no longer see that the metaphor was just a tool with inherent limitatation and boundaries.
Because the field of artificial "intelligence" is constructed around the idea of applying psychological metaphors to computational systems (a very powerful idea!) it's almost a worst case scenario for these people.
Suddenly, they're reversing the metaphors and applying computational schema to psychological processes ("aren't we really just stochastic parrots ourselves?!"); or, like here, they find themselves surprised and confused when they stumble across the natural boundaries of the metaphor experimentally.
It's because they never had sight of the boundaries in the first place and maybe never can quite see them. The words only make sense to them as literal equivalence, and so their surprise when they run into stuff like this is earnest and deep.
I don't think the boundary between "generalization" and "metaphor" is very well defined. When you go from an exemplar of 1 to 2, you're going to find all kinds of edge cases where attributes of the thing being demonstrated that had seemed to be essential turn out to not be necessary.
I think you certainly could look at LLMs as "thinking" metaphorically, but I also don't think it is necessarily only a metaphor.
This paper puts words to something I’ve noticed repeatedly with LLMs, particularly Qwen3.6. When I read its reasoning, it appears to recognise the mistake and then carry on as if it hadn’t noticed it at all.
> models often determine their answers based on implicit biases tied to question templates, then construct reasoning chains to justify their predetermined conclusions > its reasoning was correct right until the final step (Yes/No answer)
First author here, surprised to see the paper in HN! :)
When doing the paper we noticed that models are very good at generating post-hoc plausible CoT, which to me knowledge can happen quite often with relatively easy tasks.
You might be interested in reading this other paper that came out after ours: https://arxiv.org/abs/2507.05246
Natural intelligences do this too
Must we always see this restated every time? It's getting a bit stale always seeing these kinds of comments on articles about LLM.
I agree, and I very strongly dislike it, to be polite about it. It contributes absolutely nothing and is an excellent way of hand-waving away literally anything an AI model does. Saying "well people do this too" is a great way to rationalize away anything you can imagine that an AI model would be capable of, because "humans do it too so what's the big deal, guys?"
I can just immagine the response to a headline "LLM chooses mass death: thousands killed in horrific AI accident" being something like "lots of humans have caused mass death too."
It's the grounded portion of a feedback loop searching for the 'why is this happening' thinking. I'd imagine most of my own comments in this area boil down to "GIGO" most of the time.
It's a relevant comment in this instance because we're discussing concepts you need to be both trained and practiced in to reason about, and that our discipline has traditionally been blind to. Plenty of people working with LLM context issues who've never been exposed to the idea of 'subtext' or could tell you why it would matter to their direction of effort.
continuing with the "we need open training data" thread
how much of the training data had thinking traces that dont make sense to people as being actually a description of why the output should be that way?
or rather, its not productive to "we should do better with artificial intelligence"
You think, "this problem" is qn LLM problem?
No they don't, human intelligence has the ability to form an internal model of itself, which allows it to "notice" its own mistakes and change.
Many who watched the last decade knows just because its possible to noticed mistakes and change, its clearly not a reliable process.
The current political climate is well reasoned and intentional. It might not be yours or mine, however the system is working exactly as the ones paying for it have intended.
My brother and I have been arguing about that all of our lives. He believes everything is intentional and it's just a matter of discovering who benefits. I see chaos that nobody intends or controls. His political landscape is a tapestry of conspiracy theories and mine is a fog of war. I think his is more comforting, since it admits a possibility of a rational, predictable world.
I'd synthesize those as: "There is a lot of chaos with no central plan, but every small piece happens because someone believes they will benefit."
In other words, a lot of this depends on what scale/scope is being inspected. On the high level, the world is chaos rather than a meticulous and inscrutable plan of the Illuinati Shadow Cabal. On the low level, people do things for reasons, even if they're dumb ones.
With respect to the "current political climate", I'd like to suggest that a lot of dumb or "against their own interests" stuff is due to people prioritizing costly in-group loyalty signals. Their tribal interest in staying good with the club is higher than their interest against a dumb national policy.
It’s intentional in the sense that actors are acting intentionally for their own benefit (or at least what they believe is beneficial) and following incentives. Not that there is a master planner who manipulates everything
The current admin seems to have quite a few long term plans they have been working towards.
Project 2025, Maralago accords.
So far the only major policy item the Trump admin seems to have not intended was the Iran War. Israel killing the intended replacement, Iran leveraging the straight of Hormuz, and dropping three Tomahawks on an elementary school really botched that one.
Thank you for contributing absolutely nothing to this discussion.
You might not've noticed but we aren't talking about natural intelligence.
Yeah and it's not great then either
I see this when just using some chat bot that shows the “reasoning” steps (ad-hoc observation of course, it’s really cool that people are actually studying it).
It is annoying when the bot seems be “reasoning” correctly and then makes an obvious mistake at the end. And perplexing when it seems to be completely wrong and then pull the right answer out of a magic hat at the end.
I guess it makes sense; the “reasoning” steps aren’t actually doing logic, just adding more context to influence the final generation, right? But it is weird to see.
This paper made me wonder not whether the chain we can read is actual “thought,” but what conclusions we can draw by observing it. It is an output channel, but not direct access to the black box that creates it. The question pairs present an interesting experiment, but they also got me thinking about semantics: the same underlying relation can have many valid representations, and how models reason across those representations can tell us more about their stability and correctness. Basically, are models semantically consistent when given different representations of the same underlying relation? Do their answers transform as the relationship requires, and do their explanations remain consistent with that relationship?
First author here, happy to answer questions
I thought this was already widely known?
From March last year: https://transformer-circuits.pub/2025/attribution-graphs/bio...
There's no reason to believe the model's self-reported "thinking" bears any relation to the mechanics by which it arrived at some output.
It's... complicated. Yes, RL reward hacking makes it learn "bird language" and yes, reasoning traces can be misleading. However they also pretty clearly steer the final reply and not simply justify it, and can stay somewhat coherent and relevant with readability SFT and rewards. All these phenomenas coexist, they aren't mutually exclusive. Reasoning traces are still useful for debugging.
that sounds testable - if you skip the reasoning tokens, do you get the same result?
if not, then there's certainly some bearing, but not necessarily in how we read the tokens as text
It is and has been - skipping the tokens causes performance to drop. But replacing the tokens with filler causes the performance to drop, but by a lot less. You can also train models to emit broken or unrelated thinking tokens and their performance is also not much worse than the ones that are trained to output somewhat coherent thinking traces.
This points to a hypothesis that the content of the tokens is only slightly related to the mechanism by which it improves performance, and that primarily the extra tokens allow the original prompt to be processed more deeply by the model, because earlier tokens will essentially pass through the model many more times than later ones.
"Study: Communing With The Gods of Mount Olympus Via the Oracle at Delphi Is Not Always Faithful"