The takeaway for that Pfau et al. paper is slightly more nuanced than that: It can only solve a subclass of problems without CoT, and that subclass can be equivalently solved with a larger model _without_ '...'
But arguably, a larger model will not need the chain of thought a smaller model does, which means simply by scaling we're already reducing CoT.
If the people who were relying on CoT are panicking now, they should've been panicking when perceptrons became multi-layer perceptrons.
As for how universal transformers (looping transformers, but everyone has since forgotten prior work) will affect this, Will Merrill (again) has a paper here (https://arxiv.org/abs/2503.03961) that discusses exactly this.
The original universal transformers is called "universal" because if you allow for per-token looping decisions, it can theoretically be Turing complete without needing CoT (some nuance here about levels of precision used).
As for whether having little or no CoT is "unsafe": It isn't clear that the model's CoT reveal how they actually arrive at the answer. As an example, what if they provide an answer before the CoT? (https://arxiv.org/html/2603.01437v2)
If this is already in question, we shouldn't be relying on the CoT for monitoring the model's reasoning.
As always there is a lot of nuance to the topic once you get your hands dirty with the details.
>The original universal transformers is called "universal" because if you allow for per-token looping decisions, it can theoretically be Turing complete without needing CoT (some nuance here about levels of precision used).
Looping the transformer is just as turing complete as CoT. It doesn't fundamentally grant it any new theoretical capabilities. You could just scale the model into infinity with infinite context window.
Turing completeness doesn't care about the efficiency of the underlying implementation, which is fine in theoretical computer science, but if you have a model with a finite computational budget, you do actually care about the differences between write only tape vs read-write tape and single tape vs two tape. Having a fixed number of registers like a CPU also helps with reducing the number of redundant operations.
We see none of that with looped transformers, maybe we do see a fixed number of registers.
In the limited cases (below Turing completeness) there are properties of what can be done with O(log N) depth vs O(N) CoT (regular languages), if you look at the 2nd Will Merrill paper I referenced.
That analysis provided a very non-abrasive wording of their evaluation of HRM and its contributions. The comparison with a recursive / universal transformer on the same settings is telling.
"These results suggest that the performance on ARC-AGI is not an effect of the HRM architecture. While it does provide a small benefit, a replacement baseline transformer in the HRM training pipeline achieves comparable performance."
With the same data augmentation / 'test time training' setting, the vanilla Transformers do pretty well, close to the "breakthrough" HRM reported. From a brief skim, this paper is using similar settings to compare itself on ARC-AGI.
I too, want to believe in smaller models with excellent reasoning performance. But first understand what ARC-AGI tests for, what the general setting is -- the one that commercial LLMs use to compare against each other -- and what the specialised setting HRM and this paper uses as evaluation.
The naming of that benchmark lends itself to hype, as we've seen in both HRM and this paper.
The TRM paper addresses this blog post. I don't think you need to read the HRM analysis very carefully, the TRM has the advantage of being disentangled compared to the HRM, making ablations easier. I think the real value of the arcprize HRM blog post is to highlight the importance of ablation testing.
I think ARC-AGI was supposed to be a challenge for any model. The assumption being that you'd need the reasoning abilities of large language models to solve it. It turns out that this assumption is somewhat wrong. Do you mean that HRM and TRM are specifically trained on a small dataset of ARC-AGI samples, while LLMs are not? Or which difference exactly do hint at?
> Do you mean that HRM and TRM are specifically trained on a small dataset of ARC-AGI samples, while LLMs are not? Or which difference exactly do hint at?
Yes, precisely this. The question is really what is ARC-AGI evaluating for?
1. If the goal is to see if models can generalise to the ARC-AGI evals, then models being evaluated on it should not be trained on the tasks. Especially IF ARC-AGI evaluations are constructed to be OOD from the ARC-AGI training data. I don't know if they are. Further, there seems to be usage of the few-shot examples in the evals to construct more training data in the HRM case. TRM may do this via the training data via other means.
2. If the goal is that even _having seen_ the training examples, and creating more training examples (after having peeked at the test set), these evaluations should still be difficult, then the ablations show that you can get pretty far without universal/recurrent Transformers.
If 1, then I think the ARC-prize organisers should have better rules laid out for the challenge. From the blog post, I do wonder how far people will push the boundary (how much can I look at the test data to 'augment' my training data?) before the organisers say "This is explicitly not allowed for this challenge."
If 2, the organisers of the challenge should have evaluated how much of a challenge it would actually have been allowing extreme 'data augmentation', and maybe realised it wasn't that much of a challenge to begin with.
I tend to agree that, given the outcome of both the HRM and this paper, is that the ARC-AGI folks do seem to allow this setting, _and_ that the task isn't as "AGI complete" as it sets out to be.
I should probably also add: It's long been known that Universal / Recursive Transformers are able to solve _simple_ synthetic tasks that vanilla transformers cannot.
The challenge is actually scaling them up to be useful as LLMs as well (I describe why it's a challenge in the SUT paper).
It's hard to say with the way ARC-AGI is allowed to be evaluated if this is actually what is at play. My gut tells me, given the type of data that's been allowed in the training set, that some leakage of the evaluation has happened in both HRM and TRM.
But because as a field we've given up on actually carefully ensuring training and test don't contaminate, we just decide it's fine and the effect is minimal. Especially considering LLMs, the test set example leaking into the dataset is merely a drop in the bucket (I don't believe we should be dismissing it this way, but that's a whole 'nother conversation).
With these models that are challenge-targeted, it becomes a much larger proportion of what influences the model behaviour, especially if the open evaluation sets are there for everyone to look at and simply generate more. Now we don't know if we're generalising or memorising.
I think that the best way to address this potential ARC overfitting, would be to create more benchmarks - that are similar in concept, focusing on fluid intelligence, but from another angle than ARC.
Of course it is quite costly and also requires some "marketing" to actually get it established.
This would not help if no proper constraints are established on what data can and cannot be trained on. And maybe just figuring out what the goal of the benchmark is.
If it is to test generalisation capability, then what data the model being evaluated is trained on is crucial to making any conclusions.
>> Now we don't know if we're generalising or memorising.
"Now" starts around 1980 I'd say. Everyone in the field tweaks their models until they perform well on the "held-out" test set, so any ability to estimate generalisation from test-set performance goes out the window. The standard 80/20 train/test split makes it even worse.
I personally find it kind of scandalous that nobody wants to admit this in the field and yet many people are happy to make big claims about generalisation, like e.g. the "mystery" of generalising overparameterised neural nets.
You can have benchmarks with specifically constructed train-test splits for task-specific models. Train only on the train, then your results on test should be what is reported.
You can still game those benchmarks (tune your hyperparameters after looking at test results), but that setting measures for generalisation on the test set _given_ the training set specified. Using any additional data should be going against the benchmark rules, and should not be compared on the same lines.
What I'm pointing out above is that everyone games the benchmarks in the way that you say, by tuning their models until they do well on the test set. They train, they test, and they iterate until they get it right. At that point any results are meaningless for the purpose of estimating generalisation because models are effectively overfit to the test set, without ever having to train on the test set directly.
And this is standard practice, like everyone does it all the time and I believe a sizeable majority of researchers don't even understand that what they do is pointless because that's what they've been taught to do, by looking at each other's work and from what their supervisors tell them to do etc.
Btw, we don't really care about generalisation on the test set, per se. The point of testing on a held-out test set is that it's supposed to give you an estimate of a model's generalisation on truly unseen data, i.e. data that was not available to the researchers during training. That's the generalisation we're really interested in. And the reason we're interested in that is that if we deploy a model in a real-world situation (rare as that may be) it will have to deal with unseen data, not with the training data, nor with the test data.
> Now we don't know if we're generalising or memorising.
The Arc HRM blog post says:
> [...] we set out to verify HRM performance against the ARC-AGI-1 Semi-Private dataset - a hidden, hold-out set of ARC tasks used to verify that solutions are not overfit [...] 32% on ARC-AGI-1 is an impressive score with such a small model. A small drop from HRM's claimed Public Evaluation score (41%) to Semi-Private is expected. ARC-AGI-1's Public and Semi-Private sets have not been difficulty calibrated. The observed drop (-9pp) is on the high side of normal variation. If the model had been overfit to the Public set, Semi-Private performance could have collapsed (e.g., ~10% or less). This was not observed.
The question I keep coming back to is whether ARC-AGI is intended to evaluate generalisation to the task at hand. This would then mean that the test data has a meaningful distribution shift from the training data, and only a model that can perform said generalisation can do well.
This would all go out the window if the model being evaluated can _see_ the type of distribution shift it would encounter during test time. And it's unclear whether the shift is the same in the hidden set.
There are questions about the evaluations that arise from the large model performance against the smaller models, especially given the ablation studies.
Are the large models trained on the same data as these tiny models? Should they be? If they shouldn't, then why are we allowing these small models access to these in their training data?
Not exactly "vanilla Transformer", but rather "a Transformer-like architecture with recurrence".
Which is still a fun idea to play around with - this approach clearly has its strengths. But it doesn't appear to be an actual "better Transformer". I don't think it deserves nearly as much hype as it gets.
Not sure if you mean in general, but I'll answer both branches of the question.
In general:
Depending on the method of compression, you can have lossy or non-lossy compression. Using 7zip on a bunch of text files can lossless-ly compress that data. Briefly, you calculate the statistics of the data you want to compress (the dictionary), and then make the commonly re-occuring chunks describable with fewer bits (encoding). The compressed file basically contains the dictionary and the encoding.
For LLMs:
There are ways to use an LLM (or any statistical model of text) to compress text data. But the techniques use similar settings as the above, with a dictionary and an encoding, with the LLM taking the function of a dictionary. When "extracting" data from the dictionary alone, you're basically sampling from the dictionary distribution.
Quantitatively, the "loss" in "lossy" being described is literally the number of bits used for the encoding.
2nd employee at Semantics3 here. Considering all the AI available today I think things like product disambiguation becomes wayyy easier. We were trying many tricks and heuristics to identify the same products across sites.
Systems might want to anticipate changes in LLM architectures (even small changes can make a big difference kernel wise), so it's good to not "bake" too much in ahead of time.
That said, at some point it just depends where the costs lie and it might make sense hiring some GPU engineers to do what they did here for whatever architecture you're optimising for.
I'm curious how the speed is achieved is this is the technique used. Generally I expected this "masked language model" technique to be far slower since the full vocab projection needs to be computed every iteration.
I always thought the eventual technique would be some form of diffusion in continuous space, then decoding into the discrete tokens.
Also I'm guessing this is a "best guess" of how Gemini Diffusion is done?
Although marketed as such, RWKV isn't really an RNN.
In the recent RWKV7 incarnation, you could argue it's a type of Linear RNN, but past versions had an issue of taking its previous state from a lower layer, allowing for parallelism, but makes it closer to a convolution than a recurrent computation.
As for 1), I'd like to believe so, but it's hard to get people away from the addictive drug that is the easily parallelised transformer, 2) (actual) RNNs and attention mechanisms to me seem fairly powerful (expressivity wise) and perhaps most acceptable by the community.
Recent work by Feng et al from Bengio's lab focus on how attention can be formulated as an RNN ("Attention as RNN": https://arxiv.org/pdf/2405.13956) and how minimal versions of GRUs and LSTMs can be trained in parallel by removing some parameters ("Were RNNs All We Needed?" https://arxiv.org/pdf/2410.01201).
It's possible we start seeing more blended version of RNN/attention architecture exploring different LLM properties.
In particular, Aaren architecture in the former paper "can not only (i) be trained in parallel (like Transformers) but also (ii) be updated
efficiently with new tokens, requiring only constant memory for inferences (like
traditional RNNs)."
The formulations in attention as rnn have similar issues as rwkv. Fundamentally it's a question of what we call an RNN.
Personally I think it's important not to call some of these recent architectures RNNs because they have theoretical properties that do not match (read: they're worse) what we've "classically" called RNNs.
But arguably, a larger model will not need the chain of thought a smaller model does, which means simply by scaling we're already reducing CoT.
If the people who were relying on CoT are panicking now, they should've been panicking when perceptrons became multi-layer perceptrons.
reply