Kerem Yalçın retweeted
Replying to @HowToPrompt__
Someone already came up with a better algorithm: reddit.com/r/learnmachinelea…
Kerem Yalçın retweeted
Replying to @crazyMLguy
Thanks for sharing my research!
Here is the main post:
LLMs can read neural network weights. But do they actually understand what they're reading?
I tested this on 9B–30B language models.
The results surprised me.
A paper from @Poindexter_Labs, inspired by Weight-Space Programming:
Introducing Weight-Space Reading👇
Kerem Yalçın retweeted
Can an LLM understand what changing another model's weights will do without ever running that model?
A new experiment called Weight-Space Reading explores this fascinating possibility using open-weight Qwen and Muse models ranging from 9B to 30B parameters.
Some interesting findings:
73% balanced accuracy: Qwen 27B predicted subtle weight-edit effects using direct weight access, compared with 57.8% using text descriptions.
Self-inspection remains uncertain: No conclusive advantage was established for models reading their own weights.
False predictions: On Muse-30B, 34 standard edits and 6 no-op controls produced no detectable behavioral changes, yet agents frequently predicted changes.
The researchers have also released an evaluation harness, experimental code, and an open-source toolkit for weight-space programming.
The bigger question: Could LLMs eventually inspect, understand, and predict the behavior of neural networks directly from their parameters?
This could open an entirely new direction in model interpretability and automated weight editing.
GitHub: github.com/suayptalha/wsr
Paper: alphaxiv.org/abs/2611.weight…
Kerem Yalçın retweeted
"Weight-Space Reading"
Can an LLM predict how changing another model's weights will affect its behavior, without ever running that model?
This paper investigates Weight-Space Reading, where LLMs inspect the raw weights of 9B-30B parameter models using Python tools, then predict how small weight modifications will affect capabilities like math accuracy, reasoning length, and refusal behavior, without performing a single forward pass.
On subtle weight edits, Qwen3.8-27B achieves 73% balanced accuracy, compared to 57.8% when given only text descriptions of the edits.
However, the models also frequently predict changes that never happen. When 40 edits had no measurable effect on Muse-30B, the model incorrectly predicted changes for 85% of them.
The researchers also found no conclusive evidence that LLMs are better at reading their own weights than those of other models.
alphaxiv.org/abs/2611.weight…
Kerem Yalçın retweeted
Two high schoolers (me & @evoformer) improved Meta’s e-means clustering algorithm (which already outperforms k-means)!
We ran an evolutionary search over it and boom: the evolved algorithm beats e-means’ 8 epochs with just 1–3 epochs, with ~55% lower error on GPUs.
We named it s-means (“s” stands for some technical stuff... and also happens to be my name’s first letter 😅)
Huge thanks to the Meta researchers for such inspiring work!
Blog: suayptalha.github.io/s-means…
PyTorch + scikit-learn package (open source): github.com/suayptalha/s-mean…
Here’s a 2.5-min explainer video👇
Meta outperformed k-means with e-means.
We (as 2 high schoolers - me & @evoformer) outperformed e-means with s-means.
1–3 epochs of s-means > 8 epochs of e-means on A100, with +54–55% lower MSE across T4, L4 & A100 GPUs.
Fused Triton CUDA kernels, 2.24× faster than MiniBatchKMeans on CPU, zero extra passes.
Read the full writeup & grab the PyTorch/sklearn package 👇
suayptalha.github.io/s-means…
github.com/suayptalha/s-mean…
Kerem Yalçın retweeted
"Can an LLM read raw weights and predict what an edit will do, without running the model?"
Short answer: yes, kind of. With some very big asterisks.
The paper is currently trending on @askalphaxiv 🔥
Paper: alphaxiv.org/abs/2611.weight…
Code & eval harness are public: github.com/suayptalha/wsr
Here's a 5-minute explainer video for the paper and main post below 👇
LLMs can read neural network weights. But do they actually understand what they're reading?
I tested this on 9B–30B language models.
The results surprised me.
A paper from @Poindexter_Labs, inspired by Weight-Space Programming:
Introducing Weight-Space Reading👇
Kerem Yalçın retweeted
Meta outperformed k-means with e-means.
We (as 2 high schoolers - me & @evoformer) outperformed e-means with s-means.
1–3 epochs of s-means > 8 epochs of e-means on A100, with +54–55% lower MSE across T4, L4 & A100 GPUs.
Fused Triton CUDA kernels, 2.24× faster than MiniBatchKMeans on CPU, zero extra passes.
Read the full writeup & grab the PyTorch/sklearn package 👇
suayptalha.github.io/s-means…
github.com/suayptalha/s-mean…










