Introspective Diffusion Language Models
introspective-diffusion.github.io
introspective-diffusion.github.io
And then through a LoRA adapter, you can ground the diffuser on the base model’s distribution (essentially have it “compare” its proposals against what the base model would’ve generated), which effectively means: exact same byte-for-byte output for the same seed, just roughly twice as fast (which should improve even more for batched tasks).
I’m not an expert, more of a “practicing enthusiast,” so I might be missing something, but at first glance, this reads super exciting to me.
So let's say a draft model generates 5 tokens, all 5 of these can be verified in parallel with a single forward pass of the target model. The target model may only accept the first 4 tokens (or whatever) but as long as the 5 forward passes of the draft model + 1 prefill of the target model is faster than 4 forward passes of the target, you will have a speedup while maintaining the exact output distribution as the target.
then once successfully trained you get faster inference from just the diffusion model
It's the same reason there's a difference in speed between "prompt processing" and "generation". The former is just taking the pre-generated prompt and building the KV cache, which is parallel, not autoregressive and therefore way faster.
When you already know the tokens ahead of time you can calculate the probabilities of all tokens batched together, incurring significant bandwidth savings. This won't work if you're already compute bound so people with macs/etc. won't get as much benefits from this.
If that’s wrong I would certainly appreciate being corrected, though. But if it’s right, a 2.9x speed-up after rejected tokens, nearly for free, sounds amazing.
Let me explain what is going on here. This is basically a form of multi-token prediction. And speculative decoding in inference. See my earlier post[1] to understand what that is. TL;DR, in multi-token prediction you train separate LM heads to predict the next as well as next to next token as well as... Upto chosen next kth token. Training multiple LM heads is expensive and can be unnecessary, so what people typically do is have a common base for all the k heads, explained further in [1]. These guys do another variant.
Here is what they do mechanically, given a sequence p consisting of five tokens PE([p1, p2, p3, p4, p5]). Where PE(.) adds relative position info to each token.
1. Create an augmented sequence PE([p1 MASK MASK MASK MASK]). Do a training pass on that, with the ground truth sequence p1..5. Here it is trained to, for example, to predict p3 given p1+pos=-2 MASK+pos=-1 MASK+pos=0, loosely notating.
2. Then separately[2], train it as usual on PE([p1 p2 p3 p4 p5]).
Step (1) teaches it to do multi-token prediction, essentially the single LM head will (very very loosely speaking) condition on the position `k` of the special MASK token and "route" it to the "implicit" k'th LM head.
Step (2) teaches it to be a usual LLM and predict the next token. No MASK tokens involved.
So far, you have trained a multi-token predictor.
Now during inference
You use this for speculative decoding. You generate 5 tokens ahead at once with MASK tokens. And then you run that sequence through the LLM again. This has the same benefits as usual speculative decoding, namely that you can do matrix-matrix multiplication as opposed to matrix-vector. The former is more memory-bandwidth efficient due to higher arithmetic intensity.
here is an example,
query = ["what", "is", "2+2"]) prompt = PE([...query, MASK*5]) you run output = LLM(prompt). Say output is ["what", "is", "2+2", "it", "is", "4"]. Note that the NN is trained to predict the kth next token when faced with positionally encoded MASK tokens. So you get all 5 in one go. To be precise, it learns to predict "4" given ["what", "is", "2+2", MASK, MASK]. Since it does not need the "it" and "is" explicitly, you can do it in parallel with generating the "it" and the "is". "is" is predicted given ["what", "is", "2+2", MASK], for example, and that also doesn't depend on the explicit "it" being there, and thus can also be done in parallel with generating "it", which is just normal generating the next token given the query. And then you use this as a draft in your speculative decoding setup.
Their claim is that using a multi-token predictor this way as a draft model works really well. To be clear, this is still causal, the reason diffusion models have hype is because they are capable of global refinement. This is not. In the same thread as [1], I explain how increasing the number of MASK tokens, i.e increasing `k`, i.e the number of tokens you predict at once in your multi-token prediction setup quickly leads to poor quality. This paper agrees with that. They try out k=2,3,4,8. They see a drop in quality at 8 itself. So finally, this is 4-token-prediction with self-speculative decoding(sans LayerSkip or such), removing seemingly no existing limitation of such setups. It is definitely an interesting way to train MTP though.
[1] https://news.ycombinator.com/item?id=45221692
[2] Note that it is computationally a single forward pass. Attention masks help you fuse steps 1 and 2 into a single operation. However, you still have 2 separate loss values.
Then again, their results with it are great. It would be interesting to benchmark it against standard SD on a model that already uses MTP.
This startup seems to have been at it a while.
From our look into it - amazing speed, but challenges remain around time-to-first-token user experience and overall answer quality.
Can absolutely see this working if we can get the speed and accuracy up to that “good enough” position for cheaper models - or non-user facing async work.
One other question I’ve had is wondering if it’s possible to actually set a huge amount of text to diffuse as the output - using a larger body to mechanically force greater levels of reasoning. I’m sure there’s some incredibly interesting research taking place in the big labs on this.
However quality is really important. I tried that site and clicked one of their examples, "create a javascript animation". Fast response, but while it starts like this
``` Below is a self‑contained HTML + CSS + JavaScript example that creates a simple, smooth animation: a colorful ball bounces around the browser window while leaving a fading trail behind it.
<!DOCTYPE html> <html lang="en"> <head> <meta charset="UTF-8"> <title>JavaScript Bounce Animation</title> <style> body, html { margin: 0; padding: 0;
```
the answer then degrades to
``` radius: BALL_RADIUS, color: BALL_COLOR, traivD O] // array of previous {x,y} positions }; ```
Then more things start creeping in
``` // 3⃣ Bounce off walls if (ball.G 0 ball.radius < 0 || ball.x + ball.radius > _7{nas.width) { ball.vx *= -1; ibSl.x = Math.max(ball.radius, Math.min(ball.x, canvbbF4idth - ball.radius)); } if
```
and the more it goes on the worse it gets
``` Ho7 J3 Works 0 Atep | Description | ```
and
``` • prwrZ8}E6on 5 jdF wVuJg Ar touc> 2ysteners ,2 Ppawn \?) balls w>SFu the 8b$] cliM#]9 ```
This is for the demo on the front page, so I expect this is a pretty good outcome compared to what else you might ask.
I also asked it some technical details about how diffusion LLMs could work and it provided grammatically-correct plausible answers in a very short time (I don't know the tech to say if it's correct or not).
https://www.emergentmind.com/topics/dflash-block-diffusion-f...
There are several Mac implementations of it that show > 2x faster Qwen3.5 already.
Sadly, it does not perform at the level of e.g. Haiku 3.5 for tool calling, despite their own benchmarks claiming parity with Haiku 4.5, but it does compete with Flash Lite there too.
Anything with very targeted output, sufficient existing input and that benefits from a seamless feeling lends itself to dLLMs. Could see a place in tab-complete too, though Cursors model seems to be sufficiently low latency already.
I have an agentic benchmark and it shows Mercury 2 at 19/25 in 58 seconds and Mimo v2 Flash at 22/25 in 109 seconds
https://sql-benchmark.nicklothian.com/?highlight=xiaomi_mimo... (flip to the Cost vs Performance tab to see speed more graphically too)
https://chatjimmy.ai/ from Taalas seems down at the moment but if you really want speed.... 18,000 tps is something to experience
Kimi, Mimimax and GLM models provide far more robust coding assistance at sometimes no cost (financed via data sharing) or for very cheap. Output quality, tool calling reliability and task adherence tend to be far more reliable across all three over Mercury 2, so if you consider the time to get usable code including reviews, manual fixes, different prompting attempts, etc. end-to-end you'll be faster.
Only "coding" task I have found Mercury 2 to have a place for code generation is a browser desktop with simple generated applets. Think artefacts/canvas output but via a search field if the applet has been generated previously.
With other models, I need to hide the load behind a splash screen, but with Mercury 2 it is so fast that it can feel frictionless. The demo at this point is limited by the fact that venturing beyond a simple calculator or todo list, the output becomes unpredictable and I struggle to get Mercury 2 to rely on pre-made components, etc. to ensure consistent appearance and a11y.
Despite the benchmarks, cost and speed figure suggesting something different, I have had the best overall results with Haiku 4.5, simply because GPT-5.4-nano is still unwilling to play nice with my approach to UI components. I am currently experimenting with some routing, using different models for different complexity, then using loading spinners only for certain models, but even if that works reliably, any model that I cannot force to rely on UI components in a consistent manner isn't gonna work, so for the time being it'd just route between less expensive and more expensive Anthropic models.
Coding wise, one more exception can be in-line suggestions, though I have no way to fairly compare that because the tab models I know about (like Cursors) are not available via API, but Mercury 2 seems to perform solidly there, at least in Zed for a TS code base.
Basically, whether code or anything else, unless your task is truly latency dependent, I believe there are better options out there. If it is, Mercury 2 can enable some amazing things.
https://z-lab.ai/projects/dflash/
And DDTree?
[0] https://docs.inceptionlabs.ai/get-started/models#mercury-2
Consider that outputting two tokens at a time will be a (2-epsilon)x speedup over running one token at a time. As your block size increases, you quickly get to fast enough that it doesn't matter sooooo much whether you're doing blocks or actual all-at-once generation. What matters, then, is there quality trade-off for moving to block-mode output. And here it sounds like they've minimized that trade-off.
Maybe "thinking" will fix this aspect, but I see it as a serious shortcoming.
> 2025-04-12: Released I-DLM-8B, I-DLM-32B, and I-DLM-8B-LoRA on HuggingFace.
Is this old already? Not saying that's a bad thing, since it seems very sophisticated. Just curious if there's an update
I'm not a native English speaker and every now and then I see a comment in my mother tongue (downvoted to all hell of course). It's usually some kind of offhand remark.
> I understand it improved by 3x, but has the bottleneck shifted from Memory Bandwidth to Compute? Or is Memory Bandwidth still dominant?
But why did you post your comment in Japanese? We have so many good options for automated translation nowadays!
でも、なぜ日本語でコメントを投稿したんですか？最近は自動翻訳の良い選択肢がたくさんあるのに！
The original Japanese comment is clearly machine translated from another language to English. @Openpic is trolling.
I'd just downvote.