Twitxy Saved

Parshin Shojaee

Verified account This account is verified. Learn more

@ParshinShojaee

RS @apple | phd @vt_cs | understanding how LLMs reason and adapt to novel scenarios | opinions are my own
San Francisco, CA parshinsh.github.io Joined January 2020
1.6K Following 3.2K Followers 613 Posts

Latest posts 5

Pinned post

Parshin Shojaee Verified account @ParshinShojaee
Exciting new paper! 🚨

Why Do Reasoning Models Lose Coverage?

We revisit this open question with a data-centric lens 🔍 , and show how “forks in the road” situations in the post-training data can shape—and shrink—model coverage ⤵️
2 replies 10 reposts 83 likes 16K views View on X
Parshin Shojaee Verified account @ParshinShojaee
even if this work just show that a model can treat weights as code and debug them, that's by itself pretty interesting!!
Rikiya Takehi @rikiyatakehi Nov 2
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.

We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weight-sp…
View quoted post on X
5 replies 1 reposts 34 likes 1.6K views View on X
Parshin Shojaee Verified account @ParshinShojaee
I’ll be at @COLM_conf in SF next week, Oct 8–9!
Lately, I’m focusing on long horizon agents, context management, self-improvement and open-ended discovery. Would love to catch up with old and new friends, and hear what you’re working on!!

We’re also excited to share our paper on how post-training data shapes coverage shrinkage & model behavior. Say hi if you’re around!! 👀

📍 Thu, Oct 8 · 4:30pm · Imperial Ballroom · #58
➡️ Why Do Reasoning Models Lose Coverage? The Role of Data and Forks in the Road
Parshin Shojaee Verified account @ParshinShojaee May 21
Exciting new paper! 🚨

Why Do Reasoning Models Lose Coverage?

We revisit this open question with a data-centric lens 🔍 , and show how “forks in the road” situations in the post-training data can shape—and shrink—model coverage ⤵️
View quoted post on X