I think it’s important to note (i’m not an llm I know that phrase triggers you to assume I am) that they haven’t proven this as an inherent architectural issue, which I think would be the next step to the assertion.
do we know that they don’t and are incapable of reasoning, or do we just know that for x problems they jump to memorized solutions, is it possible to create an arrangement of weights that can genuinely reason, even if the current models don’t? That’s the big question that needs answered. It’s still possible that we just haven’t properly incentivized reason over memorization during training.
if someone can objectively answer “no” to that, the bubble collapses.
The puzzles the researchers have chosen are spatial and logical reasoning puzzles - so certainly not the natural domain of LLMs. The paper doesn’t unfortunately give a clear definition of reasoning, I think I might surmise it as “analysing a scenario and extracting rules that allow you to achieve a desired outcome”.
They also don’t provide the prompts they use - not even for the cases where they say they provide the algorithm in the prompt, which makes that aspect less convincing to me.
What I did find noteworthy was how the models were able to provide around 100 steps correctly for larger Tower of Hanoi problems, but only 4 or 5 correct steps for larger River Crossing problems. I think the River Crossing problem is like the one where you have a boatman who wants to get a fox, a chicken and a bag of rice across a river, but can only take two in his boat at one time? In any case, the researchers suggest that this could be because there will be plenty of examples of Towers of Hanoi with larger numbers of disks, while not so many examples of the River Crossing with a lot more than the typical number of items being ferried across. This being more evidence that the LLMs (and LRMs) are merely recalling examples they’ve seen, rather than genuinely working them out.
those particular models. It does not prove the architecture doesn’t allow it at all. It’s still possible that this is solvable with a different training technique, and none of those are using the right one. that’s what they need to prove wrong.
this proves the issue is widespread, not fundamental.
Is “model” not defined as architecture+weights? Those models certainly don’t share the same architecture. I might just be confused about your point though
It is, but this did not prove all architectures cannot reason, nor did it prove that all sets of weights cannot reason.
essentially they did not prove the issue is fundamental. And they have a pretty similar architecture, they’re all transformers trained in a similar way. I would not say they have different architectures.
that’s very true, I’m just saying this paper did not eliminate the possibility and is thus not as significant as it sounds. If they had accomplished that, the bubble would collapse, this will not meaningfully change anything, however.
also, it’s not as unreasonable as that because these are automatically assembled bundles of simulated neurons.
I think it’s important to note (i’m not an llm I know that phrase triggers you to assume I am) that they haven’t proven this as an inherent architectural issue, which I think would be the next step to the assertion.
do we know that they don’t and are incapable of reasoning, or do we just know that for x problems they jump to memorized solutions, is it possible to create an arrangement of weights that can genuinely reason, even if the current models don’t? That’s the big question that needs answered. It’s still possible that we just haven’t properly incentivized reason over memorization during training.
if someone can objectively answer “no” to that, the bubble collapses.
In case you haven’t seen it, the paper is here - https://machinelearning.apple.com/research/illusion-of-thinking (PDF linked on the left).
The puzzles the researchers have chosen are spatial and logical reasoning puzzles - so certainly not the natural domain of LLMs. The paper doesn’t unfortunately give a clear definition of reasoning, I think I might surmise it as “analysing a scenario and extracting rules that allow you to achieve a desired outcome”.
They also don’t provide the prompts they use - not even for the cases where they say they provide the algorithm in the prompt, which makes that aspect less convincing to me.
What I did find noteworthy was how the models were able to provide around 100 steps correctly for larger Tower of Hanoi problems, but only 4 or 5 correct steps for larger River Crossing problems. I think the River Crossing problem is like the one where you have a boatman who wants to get a fox, a chicken and a bag of rice across a river, but can only take two in his boat at one time? In any case, the researchers suggest that this could be because there will be plenty of examples of Towers of Hanoi with larger numbers of disks, while not so many examples of the River Crossing with a lot more than the typical number of items being ferried across. This being more evidence that the LLMs (and LRMs) are merely recalling examples they’ve seen, rather than genuinely working them out.
“even when we provide the algorithm in the prompt—so that the model only needs to execute the prescribed steps—performance does not improve”
That indicates that this particular model does not follow instructions, not that it is architecturally fundamentally incapable.
Not “This particular model”. Frontier LRMs s OpenAI’s o1/o3,DeepSeek-R, Claude 3.7 Sonnet Thinking, and Gemini Thinking.
The paper shows that Large Reasoning Models as defined today cannot interpret instructions. Their architecture does not allow it.
those particular models. It does not prove the architecture doesn’t allow it at all. It’s still possible that this is solvable with a different training technique, and none of those are using the right one. that’s what they need to prove wrong.
this proves the issue is widespread, not fundamental.
Is “model” not defined as architecture+weights? Those models certainly don’t share the same architecture. I might just be confused about your point though
It is, but this did not prove all architectures cannot reason, nor did it prove that all sets of weights cannot reason.
essentially they did not prove the issue is fundamental. And they have a pretty similar architecture, they’re all transformers trained in a similar way. I would not say they have different architectures.
Ah, gotcha
The architecture of these LRMs may make monkeys fly out of my butt. It hasn’t been proven that the architecture doesn’t allow it.
You are asking to prove a negative. The onus is to show that the architecture can reason. Not to prove that it can’t.
that’s very true, I’m just saying this paper did not eliminate the possibility and is thus not as significant as it sounds. If they had accomplished that, the bubble would collapse, this will not meaningfully change anything, however.
also, it’s not as unreasonable as that because these are automatically assembled bundles of simulated neurons.
This paper does provide a solid proof by counterexample of reasoning not occuring (following an algorithm) when it should.
The paper doesn’t need to prove that reasoning never has or will occur. It’s only demonstrates that current claims of AI reasoning are overhyped.
It does need to do that to meaningfully change anything, however.