Bluesky Thread

Self-Improving Transformers

View original thread
Self-Improving Transformers

They found that you can train LLMs on their own outputs by

1. generating *slightly harder* problems each time
2. filtering low quality via majority voting

I have thoughts. I’m going to write a blog. tl;dr it’s interesting, not the singularity

arxiv.org/abs/2502.01612
arxiv.org
Self-Improving Transformers Overcome Easy-to-Hard and Length Generalization Challenges
Large language models often struggle with length generalization and solving complex problem instances beyond their training distribution. We present a self-improvement approach where models iterativel...
51 5
1 hour later
here bsky.app/profile/timk...
Tim Kellogg @timkellogg.me
Self-Improving Transformers

They found that you can train LLMs on their own outputs by

1. generating *slightly harder* problems each time
2. filtering low quality via majority voting

Is this the singularity? Maybe, but I think it might just be benchmark saturation.

timkellogg.me/blog/2025/02...
A close-up photograph of a transparent glass coffee cup tipped over, spilling dark roasted coffee beans onto a black surface. The beans are glossy and rich brown, scattered across the frame, with some piled up inside the cup. The background is blurred, emphasizing the texture and detail of the coffee beans. The scene has warm, earthy tones with a contrast between the dark beans and the reflective surface of the glass.
2
51 likes 5 reposts

More like this

LLMs can report their own experience

there is legions of misunderstandings around LLMs

As LLMs Improve, People Adapt Their Prompts
×