Hacker News Viewer

Introspective Diffusion Language Models

by zagwdt on 4/14/2026, 7:57:33 AM

https://introspective-diffusion.github.io/

Comments

by: thepasch

If I’m reading this right, this is pretty wild. They turned a Qwen autoregressor into a diffuser by using a bunch of really clever techniques, and they vastly outperform any “native diffuser,” actually being competitive with the base model they were trained from. The obvious upside here is the <i>massive</i> speedup in generation.<p>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: <i>exact</i> same byte-for-byte output for the same seed, just roughly twice as fast (which should improve even more for batched tasks).<p>I’m not an <i>expert</i>, more of a “practicing enthusiast,” so I might be missing something, but at first glance, this reads super exciting to me.

4/14/2026, 9:46:52 AM


by: andsoitis

Is anyone here experimenting seriously with Diffusion for text generation? I’d love to learn about your experiences!

4/14/2026, 8:12:15 AM


by: ramon156

&gt; 2025-04-12: Initial code release with training and inference support.<p>&gt; 2025-04-12: Released I-DLM-8B, I-DLM-32B, and I-DLM-8B-LoRA on HuggingFace.<p>Is this old already? Not saying that&#x27;s a bad thing, since it seems very sophisticated. Just curious if there&#x27;s an update

4/14/2026, 10:09:32 AM


by: scotty79

So can you just use this and have a faster Qwen32b?<p><a href="https:&#x2F;&#x2F;huggingface.co&#x2F;yifanyu&#x2F;I-DLM-32B&#x2F;tree&#x2F;main" rel="nofollow">https:&#x2F;&#x2F;huggingface.co&#x2F;yifanyu&#x2F;I-DLM-32B&#x2F;tree&#x2F;main</a>

4/14/2026, 11:48:54 AM


by: simianwords

Can diffusion models have reasoning steps where they generate a block, introspect and then generate another until the output is satisfactory?

4/14/2026, 10:20:39 AM


by: akcd

[dead]

4/14/2026, 9:56:06 AM