Based on a quick first skim of the abstract and the introduction, the results from hierarchical reasoning (HRM) models look incredible:
> Using only 1,000 input-output examples, without pre-training or CoT supervision, HRM learns to solve problems that are intractable for even the most advanced LLMs. For example, it achieves near-perfect accuracy in complex Sudoku puzzles (Sudoku-Extreme Full) and optimal pathfinding in 30x30 mazes, where state-of-the-art CoT methods completely fail (0% accuracy). In the Abstraction and Reasoning Corpus (ARC) AGI Challenge 27,28,29 - a benchmark of inductive reasoning - HRM, trained from scratch with only the official dataset (~1000 examples), with only 27M parameters and a 30x30 grid context (900 tokens), achieves a performance
of 40.3%, which substantially surpasses leading CoT-based models like o3-mini-high (34.5%) and Claude 3.7 8K context (21.2%), despite their considerably larger parameter sizes and context lengths, as shown in Figure 1.
I'm going to read this carefully, in its entirety.
Thank you for sharing it on HN!
diwank 7 hours ago [-]
Exactly!
> It uses two interdependent recurrent modules: a *high-level module* for abstract, slow planning and a *low-level module* for rapid, detailed computations. This structure enables HRM to achieve significant computational depth while maintaining training stability and efficiency, even with minimal parameters (27 million) and small datasets (~1,000 examples).
> HRM outperforms state-of-the-art CoT models on challenging benchmarks like Sudoku-Extreme, Maze-Hard, and the Abstraction and Reasoning Corpus (ARC-AGI), where CoT methods fail entirely. For instance, it solves 96% of Sudoku puzzles and achieves 40.3% accuracy on ARC-AGI-2, surpassing larger models like Claude 3.7 and DeepSeek R1.
Erm what? How? Needs a computer and sitting down.
cs702 3 hours ago [-]
Yeah, that was pretty much my reaction. I will need time on a computer too.
I love it when authors publish working code. It's usually a good sign. If the code does what the authors claim, no one can argue with it!
diwank 2 hours ago [-]
Same! Guan’s work on sample packing during finetuning has become a staple. His openchat code is also super simple and easy to understand.
mkagenius 5 hours ago [-]
Is it talking about fine tuning existing models with 1000 examples to beat them in those tasks?
JonathanRaines 5 hours ago [-]
I advise scepticism.
This work does have some very interesting ideas, specifically avoiding the costs of backpropagation through time.
However, it does not appear to have been peer reviewed.
The results section is odd. It does not include include details of how they performed the assesments, and the only numerical values are in the figure on the front page. The results for ARC2 are (contrary to that figure) not top of the leaderboard (currently 19% compared to HRMs 5% https://www.kaggle.com/competitions/arc-prize-2025/leaderboa...)
In fields like AI/ML, I'll take a preprint with working code over peer-reviewed work without any code, always, even when the preprint isn't well edited.
Everyone everywhere can review a preprint and its published code, instead of a tiny number of hand-chosen reviewers who are often overworked, underpaid, and on tight schedules.
If the authors' claims hold up, the work will gain recognition. If the claims don't hold up, the work will eventually be ignored. Credentials are basically irrelevant.
Think of it as open-source, distributed, global review. It may be messy and ad-hoc, since no one is in charge, but it works much better than traditional peer review!
smokel 2 hours ago [-]
I sympathize partially with your views, but how would this work in practice? Where would the review comments be stored? Is one supposed to browse Hacker News to check the validity of a paper?
If a professional reviewer spots a serious problem, the paper will not make it to a conference or journal, saving us a lot of trouble.
yorwba 45 minutes ago [-]
Peer review is a way to distribute the work of identifying which papers are potentially worth reading. If you're starting from an individual paper and then ask yourself whether it was peer reviewed or not, you're doing it wrong. If you really need to know, read it yourself and accept that you might just be wasting your time.
If want to mostly read papers that have already been reviewed, start with people or organizations you trust to review papers in an area you're interested in and read what they recommend. That could be on a personal blog or through publishing a traditional journal, the difference doesn't matter much.
belter 2 hours ago [-]
> If a professional reviewer spots a serious problem
Did that ever happen? :-)
atq2119 1 hours ago [-]
Of course. As usual, you tend to not hear about it when a system we rely on works well.
hodgehog11 2 hours ago [-]
Scepticism is generally always a good idea with ML papers. Once you start publishing regularly in ML conferences, you understand that there is no traditional form of peer review anymore in this domain. The volume of papers has meant that 'peers' are often students coming to grips with parts of the field that rarely align with what they are asked to review. Conference peer review has become a 'vibe check' more than anything.
Real peer review is when other experts independently verify your claims in the arXiv submission through implementation and (hopefully) cite you in their followup work. This thread is real peer review.
rapatel0 1 hours ago [-]
THIS is so true but also not limited to ML.
Having been both a publisher and reviewer across multiple engineering, science, and bio-medical disciplines this occurs across academia.
dleeftink 2 hours ago [-]
I appreciate this insight, makes you wonder, why even publish a paper if it only amounts to a vibe check? If it's just the code we need we can get that peer reviewed through other channels.
thfuran 1 hours ago [-]
Because publications is the number that academics have to make go up.
hodgehog11 1 hours ago [-]
This and the exposure. There are so many papers on arXiv now that people often look to conference or journal publication lists.
d4rkn0d3z 4 hours ago [-]
Skepticism is best expressed by repeating the experiment and comparing results. I'm game and I have 10 days off work next month. I wonder what can be had in terms of full source and data, etc. from the authors?
Nice! They provide trained checkpoints on their GitHub. Repeating their results would be a good start.
https://github.com/sapientinc/HRM
diwank 2 hours ago [-]
I think that’s too harsh a position solely for not being peer reviewed yet. Neither of yhe original mamba1 and mamba2 papers were peer reviewed. That said, strong claims warrant strong proofs, and I’m also trying to reproduce the results locally.
sigmoid10 2 hours ago [-]
Skepticism is an understatement. There are tons of issues with this paper. Why are they comparing results of their expert model that was trained from scratch on a single task to general purpose reasoning models? It is well established in the literature that you can still beat general purpose LLMs in narrow domain tasks with specially trained, small models. The only comparison that would have made sense is one to vanilla transformers using the same nr of parameters and trained on the same input-output dataset. But the paper shows no such comparison. In fact, I would be surprised if it was significantly better, because such architecture improvements are usually very modest or not applicable in general. And insinuating that this is some significant development to improve general purpose AI by throwing in ARC is just straight up dishonest. I could probably cook up a neural net in pytorch in a few minutes that beats a hand-crafted single task that o3 can't solve in an hour. That doesn't mean that I made any progress towards AGI.
bubblyworld 2 hours ago [-]
Have you spent much time with the ARC-1 challenge? Their results on that are extremely compelling, showing results close to the initial competition's SOTA (as of closing anyway) with a tiny model and no hacks like data augmentation, pretraining, etc that all of the winning approaches leaned on heavily.
Your criticism makes sense for the maze solving and sudoku sets, of course, but I think it kinda misses the point (there are traditional algos that solve those just fine - it's more about the ability of neural nets to figure them out during training, and known issues with existing recurrent architectures).
Assuming this isn't fake news lol.
smokel 2 hours ago [-]
Looking at the code, there is a lot of data augmentation going on there. For the Sudoku and ARC data sets, they augment every example by a factor of 1,000.
That's fair, they are relabelling colours and rotating the boards. I meant more like mass generation of novel puzzles to try and train specific patterns. But you are right that technically there is some augmentation going on here, my bad.
smokel 2 hours ago [-]
Hm, I'm not so sure it's fair play for the Sudoku puzzle. Suggesting that the AI will understand the rules of the game with only 1,000 examples, and then adding 1,000,000 derived examples does not feel fair to me. Those extra examples leak a lot of information about the rules of the game.
I'm not too familiar with the ARC data set, so I can't comment on that.
bubblyworld 2 hours ago [-]
True, it leaks information about all the symmetries of the puzzle, but that's about it. I guess someone needs to test how much that actually helps - if I get the model running I'll give it a try!
sigmoid10 2 hours ago [-]
As the other commenter already pointed out, I'll believe it when I see it on the leaderboard. But even then it already lost twice against the winner of last year's competition, because that too was a general purpose LLM that could also do other things.
bubblyworld 2 hours ago [-]
Let's not move the goalposts here =) I don't think it's really fair to compare them directly like that. But I agree, this is triggering my "too good to be true" reflex very hard.
sigmoid10 2 hours ago [-]
If anything, they moved the goalpost closer to the starting line. I'm merely putting it back where it belongs.
frozenseven 30 minutes ago [-]
>does not appear to have been peer reviewed
Enough already. Please. The paper + code is here for everybody to read and test. Either it works or it doesn't. Either people will build upon it or they won't. I don't need to wait 20 months for 3 anonymous dudes to figure it out.
topspin 5 hours ago [-]
> "After completing the T steps, the H-module incorporates the sub-computation’s outcome (the final state L) and performs its own update. This H update establishes a fresh context for the L-module, essentially “restarting” its computational path and initiating a new convergence phase toward a different local equilibrium."
So they let the low-level RNN bottom out, evaluate the output in the high level module, and generate a new context for the low-level RNN. Rinse, repeat. The low-level RNNs are iterating backpropagation while the high-level is periodically kicking the low-level RNNs to get better outputs. Loops within loops. Composition.
Another interesting part:
> "Neuroscientific evidence shows that these cognitive modes share overlapping neural circuits, particularly within regions such as the prefrontal cortex and the default mode network. This indicates that the brain dynamically modulates the “runtime” of these circuits according to task complexity and potential rewards.
> Inspired by the above mechanism, we incorporate an adaptive halting strategy into HRM that enables `thinking, fast and slow'"
A scheduler that dynamically balances resources based on the necessary depth of reasoning and the available data.
I love how this paper cites parallels with real brains throughout. I believe AGI will be solved as the primitives we're developing are composed to extreme complexity, utilizing many cooperating, competing, communicating, concurrent, specialized "modules." It is apparent to me that human brain must have this complexity, because it's the only feasible way evolution had to achieve cognition using slow, low power tissue.
username135 3 hours ago [-]
As soon I read the hlm/llm split, it immediately reminded me of the human brain.
smokel 3 hours ago [-]
If I understand this correctly, it learns the rules of Sudoku by looking at 1,000 examples of (puzzle, solution) pairs. It is then able to solve previously unseen puzzles with 55% accuracy. If given millions of examples, it becomes almost perfect.
This is apparently without pretraining of any sort, which is kind of amazing. In contrast, systems like AlphaZero have the rules to go or chess built-in, and only learn the strategy, not the rules.
Off to their GitHub repository [1] to see this for myself.
AlphaZero may have the rules built in, but MuZero and the other follow-ups didn't. MuZero not only matched or surpassed AlphaZero, but it did so with less training, especially in the EfficientZero variant; notably also on the Atari playground.
smokel 3 hours ago [-]
Thanks for pointing that out.
To be fair, MuZero only learns a model of the rules for navigating its search tree. To make actual moves, it gets a list of valid actions from the game engine, so at that level it does not learn the rules of the game.
(HRM possibly does the same, and could be in the same realm as MuZero. It probably makes a lot of illegal moves.)
gavmor 2 hours ago [-]
This is "The Bitter Lesson" of AI, no? "More compute beats clever algorithm."
smokel 2 hours ago [-]
To follow up, after experimenting a bit with the source code:
1. Please, for the love of God, and for scientific reproducibility, specify library versions explicitly, and use pyproject.toml instead of an incomplete requirements.txt.
2. The 1,000 Sudoku examples are augmented with hand-coded permutation algorithms, so the actual input data set is more like 1,000,000 examples, not 1,000.
rudedogg 1 hours ago [-]
Do you have a fork or the changes? I might take a look, and python dependency hell on Sunday is no good
lispitillo 7 hours ago [-]
I hope/fear this HRM model is going to be merged with MoE very soon. Given the huge economic pressure to develop powerful LLMs I think this can be done in just a month.
The paper seems to only study problems like sudoku solving, and not question answering or other applications of LLMs. Furthermore they omit a section for future applications or fusion with current LLMs.
I think anyone working in this field can envision their applications, but the details to have a MoE with an HRM model could be their next paper.
I only skimmed the paper and I am not an expert, sure other will/can explain why they don't discuss such a new structure. Anyway, my post is just blissful ignorance over the complexity involved and the impossible task to predict change.
Edit: A more general idea is that Mixture of Expert is related to cluster of concepts and now we would have to consider a cluster of concepts related by the time they take to be grasped, so in a sense the model would have in latent space an estimation of the depth, number of layers, and time required for each concept, just like we adapt our reading style for a dense math book different to a newspaper short story.
yorwba 5 hours ago [-]
This HRM is essentially purpose-designed for solving puzzles with a small number of rules interacting in complex ways. Because the number of rules is small, a small model can learn them. Because the model is small, it can be run many times in a loop to resolve all interactions.
In contrast, language modeling requires storing a large number of arbitrary phrases and their relation to each other, so I don't think you could ever get away with a similarly small model. Fortunately, a comparatively small number of steps typically seems to be enough to get decent results.
But if you tried to use an LLM-sized model in an HRM-style loop, it would be dog slow, so I don't expect anyone to try it anytime soon. Certainly not within a month.
Maybe you could have a hybrid where an LLM has a smaller HRM bolted on to solve the occasional constraint-satisfaction task.
energy123 5 hours ago [-]
What about many small HRM models that solve conceptually distinct subtasks as determined and routed to by a master model who then analyzes and aggregates the outputs, with all of that learned during training.
buster 6 hours ago [-]
must say I am suspicious in this regard, as they don't show applications other than a Sudoku solver and don't discuss downsides.
Oras 6 hours ago [-]
and the training was only on Sudoku. Which means they need to train a small model for every problem that currently exists.
Back to ML models?
lispitillo 4 hours ago [-]
Not only on Sudoku, there is also maze solving and ARC-AGI.
OgsyedIE 5 hours ago [-]
Skimming this, there is no reason why a MoE LLM system (whether autoregressive, diffusion, energy-based or mixed) couldn't be given a nested architecture that duplicates the layout of a HRM. Combining these in different ways should allow for some novel benchmarks around efficiency and quality, which will be interesting.
3 hours ago [-]
0x000xca0xfe 5 hours ago [-]
Goodbye captchas I guess? Somehow they are still around.
belter 2 hours ago [-]
Is this not a variation of ReAct + Chain-of-Thought + Structured Planning? Or is that too unfair to the authors work?
> Using only 1,000 input-output examples, without pre-training or CoT supervision, HRM learns to solve problems that are intractable for even the most advanced LLMs. For example, it achieves near-perfect accuracy in complex Sudoku puzzles (Sudoku-Extreme Full) and optimal pathfinding in 30x30 mazes, where state-of-the-art CoT methods completely fail (0% accuracy). In the Abstraction and Reasoning Corpus (ARC) AGI Challenge 27,28,29 - a benchmark of inductive reasoning - HRM, trained from scratch with only the official dataset (~1000 examples), with only 27M parameters and a 30x30 grid context (900 tokens), achieves a performance of 40.3%, which substantially surpasses leading CoT-based models like o3-mini-high (34.5%) and Claude 3.7 8K context (21.2%), despite their considerably larger parameter sizes and context lengths, as shown in Figure 1.
I'm going to read this carefully, in its entirety.
Thank you for sharing it on HN!
> It uses two interdependent recurrent modules: a *high-level module* for abstract, slow planning and a *low-level module* for rapid, detailed computations. This structure enables HRM to achieve significant computational depth while maintaining training stability and efficiency, even with minimal parameters (27 million) and small datasets (~1,000 examples).
> HRM outperforms state-of-the-art CoT models on challenging benchmarks like Sudoku-Extreme, Maze-Hard, and the Abstraction and Reasoning Corpus (ARC-AGI), where CoT methods fail entirely. For instance, it solves 96% of Sudoku puzzles and achieves 40.3% accuracy on ARC-AGI-2, surpassing larger models like Claude 3.7 and DeepSeek R1.
Erm what? How? Needs a computer and sitting down.
The repo is at https://github.com/sapientinc/HRM .
I love it when authors publish working code. It's usually a good sign. If the code does what the authors claim, no one can argue with it!
This work does have some very interesting ideas, specifically avoiding the costs of backpropagation through time.
However, it does not appear to have been peer reviewed.
The results section is odd. It does not include include details of how they performed the assesments, and the only numerical values are in the figure on the front page. The results for ARC2 are (contrary to that figure) not top of the leaderboard (currently 19% compared to HRMs 5% https://www.kaggle.com/competitions/arc-prize-2025/leaderboa...)
In fields like AI/ML, I'll take a preprint with working code over peer-reviewed work without any code, always, even when the preprint isn't well edited.
Everyone everywhere can review a preprint and its published code, instead of a tiny number of hand-chosen reviewers who are often overworked, underpaid, and on tight schedules.
If the authors' claims hold up, the work will gain recognition. If the claims don't hold up, the work will eventually be ignored. Credentials are basically irrelevant.
Think of it as open-source, distributed, global review. It may be messy and ad-hoc, since no one is in charge, but it works much better than traditional peer review!
If a professional reviewer spots a serious problem, the paper will not make it to a conference or journal, saving us a lot of trouble.
If want to mostly read papers that have already been reviewed, start with people or organizations you trust to review papers in an area you're interested in and read what they recommend. That could be on a personal blog or through publishing a traditional journal, the difference doesn't matter much.
Did that ever happen? :-)
Real peer review is when other experts independently verify your claims in the arXiv submission through implementation and (hopefully) cite you in their followup work. This thread is real peer review.
Having been both a publisher and reviewer across multiple engineering, science, and bio-medical disciplines this occurs across academia.
Your criticism makes sense for the maze solving and sudoku sets, of course, but I think it kinda misses the point (there are traditional algos that solve those just fine - it's more about the ability of neural nets to figure them out during training, and known issues with existing recurrent architectures).
Assuming this isn't fake news lol.
https://github.com/sapientinc/HRM/blob/main/dataset/build_ar...
I'm not too familiar with the ARC data set, so I can't comment on that.
Enough already. Please. The paper + code is here for everybody to read and test. Either it works or it doesn't. Either people will build upon it or they won't. I don't need to wait 20 months for 3 anonymous dudes to figure it out.
So they let the low-level RNN bottom out, evaluate the output in the high level module, and generate a new context for the low-level RNN. Rinse, repeat. The low-level RNNs are iterating backpropagation while the high-level is periodically kicking the low-level RNNs to get better outputs. Loops within loops. Composition.
Another interesting part:
> "Neuroscientific evidence shows that these cognitive modes share overlapping neural circuits, particularly within regions such as the prefrontal cortex and the default mode network. This indicates that the brain dynamically modulates the “runtime” of these circuits according to task complexity and potential rewards.
> Inspired by the above mechanism, we incorporate an adaptive halting strategy into HRM that enables `thinking, fast and slow'"
A scheduler that dynamically balances resources based on the necessary depth of reasoning and the available data.
I love how this paper cites parallels with real brains throughout. I believe AGI will be solved as the primitives we're developing are composed to extreme complexity, utilizing many cooperating, competing, communicating, concurrent, specialized "modules." It is apparent to me that human brain must have this complexity, because it's the only feasible way evolution had to achieve cognition using slow, low power tissue.
This is apparently without pretraining of any sort, which is kind of amazing. In contrast, systems like AlphaZero have the rules to go or chess built-in, and only learn the strategy, not the rules.
Off to their GitHub repository [1] to see this for myself.
[1] https://github.com/sapientinc/HRM
To be fair, MuZero only learns a model of the rules for navigating its search tree. To make actual moves, it gets a list of valid actions from the game engine, so at that level it does not learn the rules of the game.
(HRM possibly does the same, and could be in the same realm as MuZero. It probably makes a lot of illegal moves.)
1. Please, for the love of God, and for scientific reproducibility, specify library versions explicitly, and use pyproject.toml instead of an incomplete requirements.txt.
2. The 1,000 Sudoku examples are augmented with hand-coded permutation algorithms, so the actual input data set is more like 1,000,000 examples, not 1,000.
The paper seems to only study problems like sudoku solving, and not question answering or other applications of LLMs. Furthermore they omit a section for future applications or fusion with current LLMs.
I think anyone working in this field can envision their applications, but the details to have a MoE with an HRM model could be their next paper.
I only skimmed the paper and I am not an expert, sure other will/can explain why they don't discuss such a new structure. Anyway, my post is just blissful ignorance over the complexity involved and the impossible task to predict change.
Edit: A more general idea is that Mixture of Expert is related to cluster of concepts and now we would have to consider a cluster of concepts related by the time they take to be grasped, so in a sense the model would have in latent space an estimation of the depth, number of layers, and time required for each concept, just like we adapt our reading style for a dense math book different to a newspaper short story.
In contrast, language modeling requires storing a large number of arbitrary phrases and their relation to each other, so I don't think you could ever get away with a similarly small model. Fortunately, a comparatively small number of steps typically seems to be enough to get decent results.
But if you tried to use an LLM-sized model in an HRM-style loop, it would be dog slow, so I don't expect anyone to try it anytime soon. Certainly not within a month.
Maybe you could have a hybrid where an LLM has a smaller HRM bolted on to solve the occasional constraint-satisfaction task.
Back to ML models?
[1] - https://arxiv.org/abs/2210.03629