Diffusion Beats Autoregressive in Data-Constrained Settings

Posted3 months agoActive3 months ago

djoldman

72 points

14 comments

blog.ml.cmu.eduTechstory

calmmixed

Debate

60/100

Diffusion ModelsAutoregressive ModelsMachine Learning

Key topics

Diffusion Models

Autoregressive Models

Machine Learning

A recent study suggests that diffusion models outperform autoregressive models in data-constrained settings, sparking discussion on the trade-offs between compute and data in machine learning.

Snapshot generated from the HN discussion

Discussion Activity

Light discussion

First comment

Peak period

2-3h

Avg / period

2.3

Comment distribution14 data points

Loading chart...

Based on 14 loaded comments

Key moments

01Story posted
Sep 22, 2025 at 2:21 PM EDT
3 months ago
Step 01
02First comment
Sep 22, 2025 at 3:46 PM EDT
1h after posting
Step 02
03Peak activity
4 comments in 2-3h
Hottest window of the conversation
Step 03
04Latest activity
Sep 22, 2025 at 11:23 PM EDT
3 months ago
Step 04

Generating AI Summary...

Analyzing up to 500 comments to identify key contributors and discussion patterns

Discussion (14 comments)

Showing 14 comments

blurbleblurble

3 months ago

1 reply

I have a feeling this technique might make waves: https://openreview.net/forum?id=c05qIG1Z2B#discussion

tripplyons

3 months ago

2 replies

There are definitely parallels between diffusion and reasoning models, mostly being able to spend longer to get a better solution by using a more precise ODE solver for diffusion or using more tokens for reasoning.

However, due to how diffusion models are trained, they never see their own predictions as input, so they cannot learn to store information across steps. This is the complete opposite for reasoning models.

yorwba

3 months ago

1 reply

You can train a diffusion model using its own predictions as input, no problem at all.

tripplyons

3 months ago

1 reply

At that point it is not following a diffusion training objective. I am aware of papers that do this, but I have not seen one that shows it as a better pretraining objective than something like v-prediction or flow matching.

mxwsn

3 months ago

Why is not the diffusion training objective? The technique is known as self-conditioning right? Is it an issue with conditional Tweedie's?

blurbleblurble

3 months ago

I'm probably not understanding your point but did you look at the paper? This explicitly does diffusion in an autoencoded latent space of the autoregressive prediction itself. The starting point is that prediction, but depending on how much noise is used, the diffusion model itself directly contributes to the prediction process to some degree or another.

It should be trivial to make an encoder that has some memory of at least part of the prompt (say the tailing part) and do a diffusion step there too.

smokel

3 months ago

2 replies

I fail to understand why we would lack data. Sure, there is limited (historical) text, but if we just open up all available video, and send out interactive robots into the world, we'll drown in data. Then there is simulated data, and tons of sensors that can capture vast amounts of even more data.

Edit: from the source [1], this quote pretty much sums it all up: "Our 2022 paper predicted that high-quality text data would be fully used by 2024, whereas our new results indicate that might not happen until 2028."

[1] https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-...

Legend2440

3 months ago

>send out interactive robots into the world

Easier said than done.

Robotics tends to be even more data-constrained than NLP. The real world only runs at 1x speed, and if your robot breaks something it costs real money. Simulators are simplistic compared to reality and take a lot of manual effort to build.

You will always need to make efficient use of the data you have.

imtringued

3 months ago

Robotics data isn't labeled and if you build a robot, there ain't anyone who has collected data for your particular robot.

There is also the problem that on-device learning is not yet practical.

thesz

3 months ago

2 replies

  > This paper addresses the challenge by asking: how can we trade off more compute for less data?

Autoregressive models are not matched by compute and this is the major drawback.

There is evidence that training RNN models that compute several steps with same input and coefficients (but different state) lead to better performance. It was shown in a followup to [1] that performed ablation study.

[1] https://arxiv.org/abs/1611.06188

They fixed number of time steps instead of varying it, and got better results.

Unfortunately, I forgot the title of that ablation paper.

kevinwang

3 months ago

1 reply

Not sure if you meant this because it doesn't cite the paper you mention, but it's a similar work: "An Investigation of Model-Free Planning", Guez et Al. (Deepmind) 2019 https://arxiv.org/abs/1901.03559

astrange

3 months ago

Speaking of not citing, that one could go a bit further back.

https://cdn.aaai.org/AAAI/1987/AAAI87-048.pdf

imtringued

3 months ago

It has already been proven that deep equilibrium models with a single layer are equivalent to models with a finite number of layers and the converse as well. That you can get the performance of a DEQ using a finite number of layers.

The fixed point nature of DEQs means that they inherently have a concept of self assessment how close they are to the solution. If they are at the solution, they will simply stop changing it. If not, they will keep performing calculations.

robots0only

3 months ago

This paper was just too overhyped by the authors. Also, the initial evals were very limited and very strange. This blog post does a much better job at a similar observation -- goes into details and does proper evaluation (also better attribution): https://jinjieni.notion.site/Diffusion-Language-Models-are-S...

View full discussion on Hacker News

ID: 45337433Type: storyLast synced: 11/20/2025, 2:27:16 PM

Want the full context?

Jump to the original sources

Read the primary article or dive into the live Hacker News thread when you're ready.

Open link View on HN