Although this is a neat idea, the author appears to be confused about how generative modelling works.
They repeatedly claim that previous approaches rely on factorization to reduce things to guessable chunks to avoid the 'blur problem'. This is a misunderstanding. Previous approaches solve this problem by modelling a distribution as output rather than a point. Factorization is one way of representing the distribution, but the key point is that even for the small chunks we predict an output distribution, not a point estimate.
Factored models have no issue producing TV static as an image. Each individual guess is impossible, and yet they don't produce a constant grey averaged image.
Normalizing flows, though not considered particularly efficient, are not factored at all, do not proceed in small steps, and yet still have no issues with blur because they predict a distribution.
Their approach appears to be a hard version of a latent variable model. It may be a good idea, but it isn't a fundamental change in the way they suggest.
This paper shows a nice integration of older winner-take-all ideas for learning K-modal generative models (see e.g. https://arxiv.org/abs/1612.00197, https://arxiv.org/abs/2211.14286) into modern diffusion/flow pipelines. As-implemented, I think it has some downsides:
1. K-1 extra forward passes during training
2. Inaccurate sampling behavior (will sample all K modes with equal likelihood, rather than sampling them proportionally)
However, I think both of these downsides can be mitigated by adjusting the implementation a bit more (you can have the model predict K modes jointly in one forward pass, along with probabilities of each being the min-loss mode, which you can then use for properly-weighted mode sampling at inference).
That said, I'm not entirely sure if this strategy is as generally applicable as the authors are hoping. In particular:
1. For highly-conditional image generation (like modern commercial diffusion pipelines, which use a big LLM preprocessor), most of the low-frequency color/layout decisions are already made for you by the conditioning signal. The diffusion process mostly needs to generate high-frequency details, for which there are a huge number of equally-valid modes.
2. For LLMs themselves, the sequence-generation process is already factored into a discrete classification problem, and there's no mode smearing issue to fix.
In reference to the difficulties with applying this to autoregressive LLMs - I wonder if these type of hybrids might be a good candidate for this approach.
I agree, this seems like an important discovery. And it hits a sweetspot of being effective and very simple to implement. Have a look at the pseudocode comparison at the bottom of the accompanying github page: https://explorative-modeling.github.io/
Also, a very well-structured and easy to follow blog post by the author. I very much enjoyed reading it!
This algorithm: sample a bunch of latents, train the model using the one with the lowest error.
IWAE: sample a bunch of latents, weight the loss of training the model using each one by softmax(-error). For images and text where the errors have large variance, those weights become one-hot, yielding this algorithm.
Given that these were published at ICML 2025, at which the author was a top reviewer as per their own website https://alexiglad.github.io/, I wonder how influenced they were to pursue this avenue of the back of it. They do have a related works section in appendix E, but it somehow seems to miss this. Which is odd, as the paper made a splash at the time at the conference and even made it close to the top of hacker news due to the novelty.
Of course, DDNs are a fundamentally new architecture, whereas this is more a generaliseable training strategy, but there is a very similar core.
I find the presentation of this as "factoring training" very confusing.
Factoring is something we do to distributions, which we then parameterize with our neural net.
An autoregressive model pertains to a certain factorisation of a joint distribution. And if we choose this factorisation it may have consequences for our training and sampling etc.
But to say "most models on factor generation".. I dont quite get this.
This would be better work if it didnt claim it was a new pretraining axis also. Its another way of scaling compute right?
If it were a new axis, IWAE would have an equal claim to it (as others have pointed out).
GRPO increases the likelihood of samples that are better than average, not just the single best, and decreases that of samples that are worse than average. This method doesn't even involve an explicit likelihood, so it's a completely different mechanism.
A comparison with minibatch optimal transport is in appendix A.2 of the paper.
Without anchoring on the group average it's hardly Group-Relative and without likelihood there's no Policy; that doesn't leave much of the original GRPO unchanged...
Do you have a particular paper in mind that does likelihood-free best-of-K but just calls it GRPO?
Any of the Best of N papers that exploded in popularity after GRPO.
Likelihood is not fundamental to the spirit of GRPO, any exploratory mechanism would work.
That sequential LLMs have a step-wise probability is convenient but not critical to this approach (where rejection sampling is widely used in diffusion models).
The relation to current RLVR methods I think is interesting, they do discuss it a bit but I would be curious to see more about this as well. Quote from the paper:
Exploration beyond Pretraining. The mode collapse XMs address during pretraining also often
shows up in post-training, where RL fine-tuning is known to sharpen models onto a narrow set of
behaviors [76]. Recent fixes such as pass@krewards [77] and best-of-N-aware fine-tuning [78] can
be seen through our lens as Forward XM, with a verifier standing in for ground truth data. These
fixes act only during post-training, though; pretraining with exploration may yield base models that
capture more modes in the first place, leaving RL more to select among.
Every time I read something like this, I'm thinking, "The model knows where it is at all times. It knows this because it knows where it isn't. By subtracting where it is from where it isn't, or where it isn't from where it is (whichever is greater), it obtains a difference, or deviation." At least the first couple of times through it.
No mention of GANs? As far as I remember, their images felt much more like "sampled from training data" than "averaged from training data". The idea was that you train a "discriminator" that tells generated images apart from real ones, and make the "generator" try to fool it. But somehow it lost to diffusion models, and now all AI imagery looks like slop, when GAN outputs (though imperfect) didn't look nearly as slop. I don't really know what happened though.
7567566566755677654567456745674567456654664665456667 and is have an awesome family 7o54 78o23ar and a 78o23ar that as 7o54 has always made a little of 778787ik7868i7678i678867767ju6767i767678
Although this is a neat idea, the author appears to be confused about how generative modelling works.
They repeatedly claim that previous approaches rely on factorization to reduce things to guessable chunks to avoid the 'blur problem'. This is a misunderstanding. Previous approaches solve this problem by modelling a distribution as output rather than a point. Factorization is one way of representing the distribution, but the key point is that even for the small chunks we predict an output distribution, not a point estimate.
Factored models have no issue producing TV static as an image. Each individual guess is impossible, and yet they don't produce a constant grey averaged image.
Normalizing flows, though not considered particularly efficient, are not factored at all, do not proceed in small steps, and yet still have no issues with blur because they predict a distribution.
Their approach appears to be a hard version of a latent variable model. It may be a good idea, but it isn't a fundamental change in the way they suggest.
This paper shows a nice integration of older winner-take-all ideas for learning K-modal generative models (see e.g. https://arxiv.org/abs/1612.00197, https://arxiv.org/abs/2211.14286) into modern diffusion/flow pipelines. As-implemented, I think it has some downsides:
1. K-1 extra forward passes during training
2. Inaccurate sampling behavior (will sample all K modes with equal likelihood, rather than sampling them proportionally)
However, I think both of these downsides can be mitigated by adjusting the implementation a bit more (you can have the model predict K modes jointly in one forward pass, along with probabilities of each being the min-loss mode, which you can then use for properly-weighted mode sampling at inference).
That said, I'm not entirely sure if this strategy is as generally applicable as the authors are hoping. In particular:
1. For highly-conditional image generation (like modern commercial diffusion pipelines, which use a big LLM preprocessor), most of the low-frequency color/layout decisions are already made for you by the conditioning signal. The diffusion process mostly needs to generate high-frequency details, for which there are a huge number of equally-valid modes.
2. For LLMs themselves, the sequence-generation process is already factored into a discrete classification problem, and there's no mode smearing issue to fix.
This is amazing and I think will probably end up being a pretty important development.
I was just reading this great breakdown of how diffusion Gemma works: https://newsletter.maartengrootendorst.com/p/a-visual-guide-...
In reference to the difficulties with applying this to autoregressive LLMs - I wonder if these type of hybrids might be a good candidate for this approach.
I agree, this seems like an important discovery. And it hits a sweetspot of being effective and very simple to implement. Have a look at the pseudocode comparison at the bottom of the accompanying github page: https://explorative-modeling.github.io/
Also, a very well-structured and easy to follow blog post by the author. I very much enjoyed reading it!
At first glance, this work might also be relevant for quantizing diffusion transformers (something I'm working on at the moment).
The training method is very similar to Importance Weighted Autoencoders: https://arxiv.org/abs/1509.00519
Yeah. This work is over claiming the novelty quite a bit.
I immediately thought of IWAE as well.
In what way is it similar?
This algorithm: sample a bunch of latents, train the model using the one with the lowest error.
IWAE: sample a bunch of latents, weight the loss of training the model using each one by softmax(-error). For images and text where the errors have large variance, those weights become one-hot, yielding this algorithm.
This feels extremely close in nature to being a generalisation of discrete distribution networks.
Paper: https://arxiv.org/abs/2401.00036 Project page: https://discrete-distribution-networks.github.io/
Given that these were published at ICML 2025, at which the author was a top reviewer as per their own website https://alexiglad.github.io/, I wonder how influenced they were to pursue this avenue of the back of it. They do have a related works section in appendix E, but it somehow seems to miss this. Which is odd, as the paper made a splash at the time at the conference and even made it close to the top of hacker news due to the novelty.
Of course, DDNs are a fundamentally new architecture, whereas this is more a generaliseable training strategy, but there is a very similar core.
I find the presentation of this as "factoring training" very confusing.
Factoring is something we do to distributions, which we then parameterize with our neural net.
An autoregressive model pertains to a certain factorisation of a joint distribution. And if we choose this factorisation it may have consequences for our training and sampling etc.
But to say "most models on factor generation".. I dont quite get this.
This would be better work if it didnt claim it was a new pretraining axis also. Its another way of scaling compute right?
If it were a new axis, IWAE would have an equal claim to it (as others have pointed out).
If their numbers are real and hold up at scale, every image model trained before this is now obsolete.
This is just GRPO (proposed by DeepSeek), which similarly samples many plausible generations, selects the best of K, and trains that sample.
Minibatch OT in flow matching also has a very similar mechanism, where samples from a noise distribution are matched to the closest data point.
There is a lot of prior work here that the authors neglect to discuss, which portrays this work as more novel than it actually is.
GRPO increases the likelihood of samples that are better than average, not just the single best, and decreases that of samples that are worse than average. This method doesn't even involve an explicit likelihood, so it's a completely different mechanism.
A comparison with minibatch optimal transport is in appendix A.2 of the paper.
You're right about the original GRPO proposal, but there are simplified variants that do just use best of K sampling.
GRPO (or GRPO like approaches) for diffusion/flow matching similarly can be likelihood free.
Without anchoring on the group average it's hardly Group-Relative and without likelihood there's no Policy; that doesn't leave much of the original GRPO unchanged...
Do you have a particular paper in mind that does likelihood-free best-of-K but just calls it GRPO?
Any of the Best of N papers that exploded in popularity after GRPO.
Likelihood is not fundamental to the spirit of GRPO, any exploratory mechanism would work.
That sequential LLMs have a step-wise probability is convenient but not critical to this approach (where rejection sampling is widely used in diffusion models).
Do you have an example? Would love to read one.
The relation to current RLVR methods I think is interesting, they do discuss it a bit but I would be curious to see more about this as well. Quote from the paper:
Exploration beyond Pretraining. The mode collapse XMs address during pretraining also often shows up in post-training, where RL fine-tuning is known to sharpen models onto a narrow set of behaviors [76]. Recent fixes such as pass@krewards [77] and best-of-N-aware fine-tuning [78] can be seen through our lens as Forward XM, with a verifier standing in for ground truth data. These fixes act only during post-training, though; pretraining with exploration may yield base models that capture more modes in the first place, leaving RL more to select among.
That's fair.
On the other hand, Minibatch OT (optimal transport) was one of the more fundamental advancements early on in flow matching & rectified flow models.
This best of K approach effectively discards matches that would otherwise be preserved by optimal transport assignment.
I still think the authors here are too broad regarding their claimed novelty.
I don't have enough context to know if this is actually cool or not, but it seems like it!
Every time I read something like this, I'm thinking, "The model knows where it is at all times. It knows this because it knows where it isn't. By subtracting where it is from where it isn't, or where it isn't from where it is (whichever is greater), it obtains a difference, or deviation." At least the first couple of times through it.
No mention of GANs? As far as I remember, their images felt much more like "sampled from training data" than "averaged from training data". The idea was that you train a "discriminator" that tells generated images apart from real ones, and make the "generator" try to fool it. But somehow it lost to diffusion models, and now all AI imagery looks like slop, when GAN outputs (though imperfect) didn't look nearly as slop. I don't really know what happened though.
7567566566755677654567456745674567456654664665456667 and is have an awesome family 7o54 78o23ar and a 78o23ar that as 7o54 has always made a little of 778787ik7868i7678i678867767ju6767i767678
Project page: https://explorative-modeling.github.io/
Paper: https://arxiv.org/abs/2607.27372
Code: https://github.com/alexiglad/XM