Maximilian Schlegel

Emergent temporal abstractions in autoregressive models enable hierarchical reinforcement learning Large-scale autoregressive models pretrained on next-token prediction and finetuned with reinforcement learning (RL) have achieved unprecedented success on many problem domains. During RL, these model...

Well this is exciting: arxiv.org/abs/2512.20605

YouTube video by MAIN Conference Joao Sacramento - “Emergent temporal abstractions in autoregressive models enable hierarchical RL”

Check out João Sacramento’s talk about our work on “Internal Reinforcement Learning”!

Here’s the link to the recording: youtu.be/cx_MIhvAOYM?...

And here’s the link to the new paper (arxiv.org/abs/2512.20605)

Seijin Kobayashi, et al.: Emergent temporal abstractions in autoregressive models enable hierarchical reinforcement learning https://arxiv.org/abs/2512.20605 https://arxiv.org/pdf/2512.20605 https://arxiv.org/html/2512.20605

Live : Talk #3 at #MAIN2025

Joao Sacramento (Google):

“Emergent temporal abstractions in autoregressive models enable hierarchical reinforcement learning”

#RL #NeuroAI

Weight decay induces low-rank attention layers The effect of regularizers such as weight decay when training deep neural networks is not well understood. We study the influence of weight decay as well as $L2$-regularization when training neural ne...

Super cool paper

I don’t get why nobody seems to care about “Trending #”. You barely find it in these “bluesky is missing this list of features” posts but imo it was the main reason why twitter became a “news platform” and was used so much by people without a bubble

Since they now have their own track, I’d say high school students lol

isn’t the idea of it “deterrence”? hard to tell if your review will be marked as highly irresponsible, so to avoid trouble altogether, maybe more will produce better reviews

—> RNN :)

This is great, thanks!

Reconnecting with the Twitter homies on Bluesky

And tbh you rightfully did I just skimmed the article then and misread this haha

detailed results of this are available at slatestarcodex.com/Stuff/aituri.... quickly checked for the top5 pictures among those who hate ai art. The two winning ones are AI (3,4,5: human) but close in style to famous impressionist paintings as others pointed out

Wait the data is actually available lol

Yes I agree! And indeed it is very beautiful.

Yeah, sorry that’s on me, there’s only the winning one, 31: “This was the most popular picture in the dataset.” The creator gives the prompt in the comments under the article: “Impressionist painting of a countryside café terrace during twilight, 1883.”

while I agree with you that ai art can be very creative I don’t see how the test results prove these people wrong. The winning ai pictures are just very close in style to famous impressionist paintings which nobody would claim lack soul

let’s goo! gpu cluster secured

Isn’t the reason for most people’s problem with AI art the copyright stuff, ie “generate painting of <thing> in the style of <artist>” leading to deceptively real looking paintings? I don’t believe people think it’s generally slob (even though many might not admit this publicly haha)

Dear Santa,

My name is Max, I’m 23 years old and live in Zurich. For Christmas I would love to find a

24xH100 GPU cluster

under the tree!

Sincerely,
Max