> To resolve this, we propose a parallel multimodal diffusion framework, MMaDA-Parallel, that enables continuous, bidirectional interaction between text and images throughout the entire denoising trajectory.
> (ParaRL), a novel strategy that applies semantic rewards along the trajectory to enforce cross-modal consistency.
(emphasis mine)
This sounds really cool. The fact that one generation "attends" to the other is really interesting. I'm curious if this would hold for other modalities. I'm thinking coding specific applications, where things can change once something is generated. My hunch is that coding would benefit a lot from this approach, because the "manual" way of writing code often resembles diffusion more than autoregressive (that is, we often edit something here, then because we did that we have to import something, then change something there, then that leads to further changes, etc).
For now coding seems to benefit a lot from <thinking> -> <coding> -> <env_feedback> -> <reflexion> -> <thinking> -> <coding>, but this seems at a glance to be shoehorned in for autoregressive generation... GPT5 in particular seems to be better at this, with multiple "tool calls" interleaved in its thinking sessions. I wonder if this would get better with the paralel denoising thing proposed here, where both thinking and coding are done in paralel, and one can "attend" to the other. Add some feedback (linters, compilers, LSPs, tests, etc.) and this can go places. If it works.
Diffusion text models aren't new, I've made them at home. Also, plenty of frontier models are good at tool calling, GPT-5 has just been trained to do it more so that it appears to do better at coding exercises with codex/IDEs.
If you haven't tried an agentic IDE such as Cursor yet, or at least an extension such as Copilot, I would recommend checking them out and trying out Anthropic's models as well.
Do you have any examples / papers where they do the parallel thing proposed here? I've tried googles diffusion coding model, but AFAICT they don't do parallel thinking & coding. It seems to just take a prompt and output code.
What's cool with this thinking & generation in parallel is that one can attend to the other. So you're not limited by prompt influences code, but can do prompt influences both thinking and code, and code can influence thinking and thinking can influence code.
They use bidirectional attention between modalities, not within the same modality. This doesn’t change much in the context you're referring to (coding). How do you think "thinking" works in current SOTA models like GPT-5-Thinking/Pro? When generating code, the model's "thinking" already attends to the code, and both influence each other during generation. "Reasoning" models modify the code as they generate it, they delete it, revise it, and adjust their internal reasoning based on the new tokens they produce during the "thinking" process. There are dozens of denoising models created for text, they are not good at it and parallel sampling between modalities will not change that.
They cannot "edit" the code though, like you can with diffusion. They must emit all tokens again, or a patch/diff which is not directly connected to the previous stream of tokens.
LLMs can "edit" code, but as you say, they do it differently from diffusion models. They operate directly on long text sequences and use much more context, which is one reason they currently work better for coding. Diffusion models for code aren't a new idea, people have tried different designs, but so far they tend to underperform autoregressive LLMs, probably because denoising over discrete tokens is harder to make work than straightforward next token prediction.
> We provide two varients of MMaDA-Parallel with different tokenizers. MMaDA-Parallel-A is trained with tokenizer Amused-VQ, and MMaDA-Parallel-M is trained with tokenizer Magvitv2.
Out of curiosity, is it possible this suffers from the same issues Anthropic found where reasoning expressed by the model and actual internal reasoning differ?
I think this is likely to happen in all models since their internal reasoning is not in the same form as the output. This is probably true also for humans.
This may solve the additional clouding that comes from LLMs using what is an effectively an iteration of instants to introspect the past. You cannot ask a autoregressive model what the thinking was behind the output because the only memory it has of the past is the output. It has to infer what it meant just the same as anyone else would.
To some extent this probably also happens in humans. You have richer memories, but you still do a lot of post hoc rationalisation.
Native latent reasoning, with latent aware RL scaffolding and all the rest will have to be built. If you use the direct text framework, you get confabulation / hallucination issues from the divergence between the tokens in the context and the rich activation representation that resulted in the output.
There are all sorts of places where the text and output is at least one degree of separation from the underlying activation vectors or other representations handled by a model, from floating point precision all the way up to tokenization abstraction, and a lot of experiments get run as if the tokens and context and representations are all one unified data concept. Have to match data abstractions appropriately, or the weird edge cases will break things in unexpected ways.
This looks awesome. Although from a UX perspective might not be as good as streaming token by token for text generation use cases. However for image gen and editing - 100%
> To resolve this, we propose a parallel multimodal diffusion framework, MMaDA-Parallel, that enables continuous, bidirectional interaction between text and images throughout the entire denoising trajectory.
> (ParaRL), a novel strategy that applies semantic rewards along the trajectory to enforce cross-modal consistency.
(emphasis mine)
This sounds really cool. The fact that one generation "attends" to the other is really interesting. I'm curious if this would hold for other modalities. I'm thinking coding specific applications, where things can change once something is generated. My hunch is that coding would benefit a lot from this approach, because the "manual" way of writing code often resembles diffusion more than autoregressive (that is, we often edit something here, then because we did that we have to import something, then change something there, then that leads to further changes, etc).
For now coding seems to benefit a lot from <thinking> -> <coding> -> <env_feedback> -> <reflexion> -> <thinking> -> <coding>, but this seems at a glance to be shoehorned in for autoregressive generation... GPT5 in particular seems to be better at this, with multiple "tool calls" interleaved in its thinking sessions. I wonder if this would get better with the paralel denoising thing proposed here, where both thinking and coding are done in paralel, and one can "attend" to the other. Add some feedback (linters, compilers, LSPs, tests, etc.) and this can go places. If it works.
Diffusion text models aren't new, I've made them at home. Also, plenty of frontier models are good at tool calling, GPT-5 has just been trained to do it more so that it appears to do better at coding exercises with codex/IDEs.
If you haven't tried an agentic IDE such as Cursor yet, or at least an extension such as Copilot, I would recommend checking them out and trying out Anthropic's models as well.
Do you have any examples / papers where they do the parallel thing proposed here? I've tried googles diffusion coding model, but AFAICT they don't do parallel thinking & coding. It seems to just take a prompt and output code.
What's cool with this thinking & generation in parallel is that one can attend to the other. So you're not limited by prompt influences code, but can do prompt influences both thinking and code, and code can influence thinking and thinking can influence code.
They use bidirectional attention between modalities, not within the same modality. This doesn’t change much in the context you're referring to (coding). How do you think "thinking" works in current SOTA models like GPT-5-Thinking/Pro? When generating code, the model's "thinking" already attends to the code, and both influence each other during generation. "Reasoning" models modify the code as they generate it, they delete it, revise it, and adjust their internal reasoning based on the new tokens they produce during the "thinking" process. There are dozens of denoising models created for text, they are not good at it and parallel sampling between modalities will not change that.
They cannot "edit" the code though, like you can with diffusion. They must emit all tokens again, or a patch/diff which is not directly connected to the previous stream of tokens.
LLMs can "edit" code, but as you say, they do it differently from diffusion models. They operate directly on long text sequences and use much more context, which is one reason they currently work better for coding. Diffusion models for code aren't a new idea, people have tried different designs, but so far they tend to underperform autoregressive LLMs, probably because denoising over discrete tokens is harder to make work than straightforward next token prediction.
Be aware that the project page has the wrong Arxiv link at the time of writing. This is the correct one:
https://arxiv.org/abs/2511.09611
Interesting approach and a very readable paper.
> We provide two varients of MMaDA-Parallel with different tokenizers. MMaDA-Parallel-A is trained with tokenizer Amused-VQ, and MMaDA-Parallel-M is trained with tokenizer Magvitv2.
tyfeld/MMaDA-Parallel-A: https://huggingface.co/tyfeld/MMaDA-Parallel-A/tree/main
tyfeld/MMaDA-Parallel-M: https://huggingface.co/tyfeld/MMaDA-Parallel-M/tree/main
Out of curiosity, is it possible this suffers from the same issues Anthropic found where reasoning expressed by the model and actual internal reasoning differ?
I think this is likely to happen in all models since their internal reasoning is not in the same form as the output. This is probably true also for humans.
This may solve the additional clouding that comes from LLMs using what is an effectively an iteration of instants to introspect the past. You cannot ask a autoregressive model what the thinking was behind the output because the only memory it has of the past is the output. It has to infer what it meant just the same as anyone else would.
To some extent this probably also happens in humans. You have richer memories, but you still do a lot of post hoc rationalisation.
Native latent reasoning, with latent aware RL scaffolding and all the rest will have to be built. If you use the direct text framework, you get confabulation / hallucination issues from the divergence between the tokens in the context and the rich activation representation that resulted in the output.
There are all sorts of places where the text and output is at least one degree of separation from the underlying activation vectors or other representations handled by a model, from floating point precision all the way up to tokenization abstraction, and a lot of experiments get run as if the tokens and context and representations are all one unified data concept. Have to match data abstractions appropriately, or the weird edge cases will break things in unexpected ways.
This looks awesome. Although from a UX perspective might not be as good as streaming token by token for text generation use cases. However for image gen and editing - 100%
[dead]
[flagged]
[flagged]