> Without specification, we employ a decoder-only language model GPT2 (Radford et al., 2019) with a configuration of 4 layers, 32 hidden dimensions, and 4 attention
heads.
Yeah, ok. The research is interesting, warranted, but writing an article about it, and leading with the conclusions gathered from toy models and implying this generalises to production LLMs is useless.
We've been here before with small models. Training on LLM outputs leads to catastrophic collapse. Every outlet led with this. But no-one red the fine-print, they were testing on small toy models, and were using everything that came out to re-train. Of course it's gonna fail. L3 / phi / gpt-oss models showed that you can absolutely train on synthetic datasets and have great results.
Research in this area is good, and needed. Mainly to understand limitations, discover if there are any scale levels where "emergent" stuff appears and so on. But writing articles based on incipient research, based on tiny models is not worth the effort.
kazinator · 38m ago
> conclusions gathered from toy models and implying this generalises to production LLMs is useless
You are just trotting out the tired argument that model size magically fixes the issues, rather than just improves the mirage, and so nothing can be known about models with M parameters by studying models with N < M parameters.
Given enough parameters, a miraculous threshold is reached whereby LLMs switch from interpolating to extrapolating.
Sure!
ricardobeat · 23m ago
That’s what has been seen in practice though. SOTA LLMs have been shown again and again to solve problems unseen in their data set; and despite their shortcomings they have become extremely useful for a wide variety of tasks.
OtherShrezzing · 28m ago
I think it is worth writing about simply because it might get the (cost constrained) researcher’s work in front of someone who has the near-unlimited research budgets at one of the big AI companies.
willvarfar · 1h ago
Doing analysis on small models or small data is perfectly valid if the results extrapolate to large models. Which is why right now we're looking at new research papers that are still listing the same small datasets and comparing to the same small models that papers five years ago did.
NitpickLawyer · 1h ago
I have nothing against researching this, I think it's important. My main issue is with articles choosing to grab a "conclusion" and imply it extrapolates to larger models, without any support for that. They are going for the catchy title first, fine-print be damned.
willvarfar · 1h ago
I was just at the KDD conference and the general consensus agreed with this paper. There was only one keynoter who just made the assumption that LLMs are associated with reasoning, which was jarring as the previous keynoter had just explained at length why we need a neuro-symbolic approach instead.
The thing is, I think the current companies making LLMs are _not_ trying to be correct or right. They are just trying to hide it better. In the business future for AI the coding stuff that we focus on on HN - how AI can help/impact us - is just a sideline.
The huge-money business future of LLMs is to end consumers not creators and it is product and opinion placement and their path to that is to friendship. They want their assistant to be your friend, then your best friend, then your only friend, then your lover. If the last 15 years of social media has been about discord and polarisation to get engagement, the next 15 will be about friendship and love even though that leads to isolation.
None of this needs the model to grow strong reasoning skills. That's not where the real money is. And CoT - whilst super great - is just as effective if it's hiding better that its giving you the wrong answer (by being more internally consistent) than if its giving you a better answer?
mdp2021 · 55m ago
> None of this needs the model to grow strong reasoning skills. That's not where the real money is
"And the world is more and more complex, and the administrations are less and less prepared"
(~~ Henry Kissinger)
refulgentis · 1h ago
Not sure what all this is about, I somewhat regret taking a breaking from coding with LLMs to have it explained to me its all a mirage and a secret and sloppy plan for getting me an automagic egirl or something. ;)
XenophileJKO · 27m ago
Right? Oh this fairly novel solution the the problem I was having that works and is well tested. Oh throw it away.. sorry the model can't think of stuff..
Back to square one!!
kazinator · 34m ago
Because model size is a trivial parameter, and not a new paradigm.
What you're saying is like, you can't extrapolate that long division works on 100 digit numbers because you only worked through it using 7 digit numbers and a few small polynomials.
suddenlybananas · 1h ago
>Training on LLM outputs leads to catastrophic collapse. Every outlet led with this. But no-one red the fine-print, they were testing on small toy models, and were using everything that came out to re-train. Of course it's gonna fail. L3 / phi / gpt-oss models showed that you can absolutely train on synthetic datasets and have great results
You're conflating two very different things. Training on synthetic data one time is very different than cyclically training models on their own data. It has nothing to do with model size.
NitpickLawyer · 1h ago
Perhaps I worded it poorly. My main point was that articles focus on the wrong thing. Most coverage of that paper was "Using LLM generated data leads to CATASTROPHIC collapse". Without reading the fineprint.
> [...] cyclically training models on their own data. It has nothing to do with model size.
Of course it does. GRPO is basically "training models on their own data". You sample, you check for a known truth, you adapt the weights. Repeat. And before GRPO there was RLAIF which showed improving scores at 3 "stages" of generate - select - re-train. With diminishing returns after 3 stages, but no catastrophic collapse.
My main point was about articles and cherrypicking catchy phrases, not criticising research. We need the research. But we also need good articles that aren't written just for the negativity sells titles.
cheeky edit: see this thread [1]. I know slashdot has fallen a lot in the last years, but I skimmed the root comments. Not one addressing the "toy" model problem. Everyone reads the title, and reinforces their own biases. That's the main problem I was trying to address.
If you have a ground truth that you're comparing to, that's not training on your own data.
tankenmate · 1h ago
"Training on synthetic data one time is very different than cyclically training models on their own data.", but every one with even a modicum of understanding of feedback knows that cyclic training on its own output will end in tears; it's bordering on a tautologic inverse.
Frieren · 2h ago
This assessment fits with my anecdotal evidence. LLMs just cannot reason in any basic way.
LLMs have a large knowledge base that can be spit out at a moment notice. But they have zero insight on its contents, even when the information has just been asked a few lines before.
Most of the "intelligence" that LLMs show is just the ability to ask in the correct way the correct questions mirrored back to the user. That is why there is so many advice on how to do "proper prompting".
That and the fact that most questions have already been asked before as anyone that spend some time in StackOverflow back in the day realized. And memory and not reasoning is what is needed to answer them.
PeterStuer · 1h ago
Please don't tell me you were one of those marking every SO question as duplicate, more often than not missing the entire nuance in the question that made it not a duplicate at all, and the answers to the so called previously asked question utterly unusable?
This was one of those infuriating things that drove so many away from SO and jump ship the second there was an alternative.
antihipocrat · 1h ago
I'm not sure why duplicates were ever considered an issue. For certain subjects (like JS) things evolved so quickly during the height of SO that even a year old answer was outdated.
That and search engines seemed to promote more recent content.. so an old answer sank under the ocean of blog spam
ceejayoz · 2m ago
SO wanted to avoid being a raw Q&A site in favor of something more like a wiki.
If a year-old answer on a canonical question is incorrect, you edit it.
Frieren · 40m ago
I was "playing" the gamification part of StackOverflow. I wanted to ask a good question for points. But it was very difficult because any meaningful question had already been asked. It was way easier to find questions to answer.
ceejayoz · 1h ago
Every time I ask people for an example of this, and get one, I agree with the duplicate determination. Sometimes it requires a little skimming of the canonical answers past just the #1 accepted one; sometimes there's a heavily upvoted clarification in a top comment, but it's usually pretty reasonable.
mirekrusin · 2h ago
Hold on their evaluation tasks are based on rotating letters in text? Isn't this known weak area for token based models?
Terr_ · 1h ago
I think that's the point, really: It's a reliable and reproducible weakness, but also one where the model can be trained to elicit impressive-looking "reasoning" about what the problem is and how it "plans" to overcome it.
Then when it fails to apply the "reasoning", that's evidence the artificial expertise we humans perceived or inferred is actually some kind of illusion.
Kind of like a a Chinese Room scenario: If the other end appears to talk about algebra perfectly well, but just can't do it, that's evidence you might be talking to a language-lookup machine instead of one that can reason.
boredhedgehog · 37m ago
> Then when it fails to apply the "reasoning", that's evidence the artificial expertise we humans perceived or inferred is actually some kind of illusion.
That doesn't follow, if the weakness of the model manifests on a different level we wouldn't call rational in a human.
For example, a human might have dyslexia, a disorder on the perceptive level. A dyslexic can understand and explain his own limitation, but that doesn't help him overcome it.
hooskerdu · 59m ago
Reminds me of a number of grad students I knew who could “talk circles” around all sorts of subjects but failed to ever be able to apply anything.
Terr_ · 56m ago
Heh, but just because a human can fail at something doesn't mean everything that fails at it is human. :p
syllogism · 1h ago
It's interesting that there's still such a market for this sort of take.
> In a recent pre-print paper, researchers from the University of Arizona summarize this existing work as "suggest[ing] that LLMs are not principled reasoners but rather sophisticated simulators of reasoning-like text."
What does this even mean? Let's veto the word "reasoning" here and reflect.
The LLM produces a series of outputs. Each output changes the likelihood of the next output. So it's transitioning in a very large state space.
Assume there exists some states that the activations could be in that would cause the correct output to be generated. Assume also that there is some possible path of text connecting the original input to such a success state.
The reinforcement learning objective reinforces pathways that were successful during training. If there's some intermediate calculation to do or 'inference' that could be drawn, writing out a new text that makes that explicit might be a useful step. The reinforcement learning objective is supposed to encourage the model to learn such patterns.
So what does "sophisticated simulators of reasoning-like text" even mean here? The mechanism that the model uses to transition towards the answer is to generate intermediate text. What's the complaint here?
It makes the same sort of sense to talk about the model "reasoning" as it does to talk about AlphaZero "valuing material" or "fighting for the center". These are shorthands for describing patterns of behaviour, but of course the model doesn't "value" anything in a strictly human way. The chess engine usually doesn't see a full line to victory, but in the games it's played, paths which transition through states with material advantage are often good -- although it depends on other factors.
So of course the chain-of-thought transition process is brittle, and it's brittle in ways that don't match human mistakes. What does it prove that there are counter-examples with irrelevant text interposed that cause the model to produce the wrong output? It shows nothing --- it's a probabilistic process. Of course some different inputs lead to different paths being taken, which may be less successful.
bubblyworld · 21m ago
Not sure why everyone is downvoting you as I think you raise a good point - these anthropomorphic words like "reasoning" are useful as shorthands for describing patterns of behaviour, and are generally not meant to be direct comparisons to human cognition. But it goes both ways. You can still criticise the model on the grounds that what we call "reasoning" in the context of LLMs doesn't match the patterns we associate with human "reasoning" very well (such as ability to generalise to novel situations), which is what I think the authors are doing.
Gusarich · 2h ago
The article already seems outdated on the first day. The key points about SFT are irrelevant in the era of RL.
acosmism · 2h ago
remind me in 2 days
Martin_Silenus · 1h ago
If only we could train people like that to see their reasoning output...
floppiplopp · 1h ago
'Chain-of-thought AI "degrades significantly" when asked to generalize beyond training.' - yeah thanks Captain Obvious.
Yeah, ok. The research is interesting, warranted, but writing an article about it, and leading with the conclusions gathered from toy models and implying this generalises to production LLMs is useless.
We've been here before with small models. Training on LLM outputs leads to catastrophic collapse. Every outlet led with this. But no-one red the fine-print, they were testing on small toy models, and were using everything that came out to re-train. Of course it's gonna fail. L3 / phi / gpt-oss models showed that you can absolutely train on synthetic datasets and have great results.
Research in this area is good, and needed. Mainly to understand limitations, discover if there are any scale levels where "emergent" stuff appears and so on. But writing articles based on incipient research, based on tiny models is not worth the effort.
You are just trotting out the tired argument that model size magically fixes the issues, rather than just improves the mirage, and so nothing can be known about models with M parameters by studying models with N < M parameters.
Given enough parameters, a miraculous threshold is reached whereby LLMs switch from interpolating to extrapolating.
Sure!
The thing is, I think the current companies making LLMs are _not_ trying to be correct or right. They are just trying to hide it better. In the business future for AI the coding stuff that we focus on on HN - how AI can help/impact us - is just a sideline.
The huge-money business future of LLMs is to end consumers not creators and it is product and opinion placement and their path to that is to friendship. They want their assistant to be your friend, then your best friend, then your only friend, then your lover. If the last 15 years of social media has been about discord and polarisation to get engagement, the next 15 will be about friendship and love even though that leads to isolation.
None of this needs the model to grow strong reasoning skills. That's not where the real money is. And CoT - whilst super great - is just as effective if it's hiding better that its giving you the wrong answer (by being more internally consistent) than if its giving you a better answer?
"And the world is more and more complex, and the administrations are less and less prepared"
(~~ Henry Kissinger)
Back to square one!!
What you're saying is like, you can't extrapolate that long division works on 100 digit numbers because you only worked through it using 7 digit numbers and a few small polynomials.
You're conflating two very different things. Training on synthetic data one time is very different than cyclically training models on their own data. It has nothing to do with model size.
> [...] cyclically training models on their own data. It has nothing to do with model size.
Of course it does. GRPO is basically "training models on their own data". You sample, you check for a known truth, you adapt the weights. Repeat. And before GRPO there was RLAIF which showed improving scores at 3 "stages" of generate - select - re-train. With diminishing returns after 3 stages, but no catastrophic collapse.
My main point was about articles and cherrypicking catchy phrases, not criticising research. We need the research. But we also need good articles that aren't written just for the negativity sells titles.
cheeky edit: see this thread [1]. I know slashdot has fallen a lot in the last years, but I skimmed the root comments. Not one addressing the "toy" model problem. Everyone reads the title, and reinforces their own biases. That's the main problem I was trying to address.
1 - https://slashdot.org/story/25/08/11/2253229/llms-simulated-r...
LLMs have a large knowledge base that can be spit out at a moment notice. But they have zero insight on its contents, even when the information has just been asked a few lines before.
Most of the "intelligence" that LLMs show is just the ability to ask in the correct way the correct questions mirrored back to the user. That is why there is so many advice on how to do "proper prompting".
That and the fact that most questions have already been asked before as anyone that spend some time in StackOverflow back in the day realized. And memory and not reasoning is what is needed to answer them.
This was one of those infuriating things that drove so many away from SO and jump ship the second there was an alternative.
That and search engines seemed to promote more recent content.. so an old answer sank under the ocean of blog spam
If a year-old answer on a canonical question is incorrect, you edit it.
Then when it fails to apply the "reasoning", that's evidence the artificial expertise we humans perceived or inferred is actually some kind of illusion.
Kind of like a a Chinese Room scenario: If the other end appears to talk about algebra perfectly well, but just can't do it, that's evidence you might be talking to a language-lookup machine instead of one that can reason.
That doesn't follow, if the weakness of the model manifests on a different level we wouldn't call rational in a human.
For example, a human might have dyslexia, a disorder on the perceptive level. A dyslexic can understand and explain his own limitation, but that doesn't help him overcome it.
> In a recent pre-print paper, researchers from the University of Arizona summarize this existing work as "suggest[ing] that LLMs are not principled reasoners but rather sophisticated simulators of reasoning-like text."
What does this even mean? Let's veto the word "reasoning" here and reflect.
The LLM produces a series of outputs. Each output changes the likelihood of the next output. So it's transitioning in a very large state space.
Assume there exists some states that the activations could be in that would cause the correct output to be generated. Assume also that there is some possible path of text connecting the original input to such a success state.
The reinforcement learning objective reinforces pathways that were successful during training. If there's some intermediate calculation to do or 'inference' that could be drawn, writing out a new text that makes that explicit might be a useful step. The reinforcement learning objective is supposed to encourage the model to learn such patterns.
So what does "sophisticated simulators of reasoning-like text" even mean here? The mechanism that the model uses to transition towards the answer is to generate intermediate text. What's the complaint here?
It makes the same sort of sense to talk about the model "reasoning" as it does to talk about AlphaZero "valuing material" or "fighting for the center". These are shorthands for describing patterns of behaviour, but of course the model doesn't "value" anything in a strictly human way. The chess engine usually doesn't see a full line to victory, but in the games it's played, paths which transition through states with material advantage are often good -- although it depends on other factors.
So of course the chain-of-thought transition process is brittle, and it's brittle in ways that don't match human mistakes. What does it prove that there are counter-examples with irrelevant text interposed that cause the model to produce the wrong output? It shows nothing --- it's a probabilistic process. Of course some different inputs lead to different paths being taken, which may be less successful.