Vedang Lad
@vedanglad
Followers
511
Following
480
Statuses
118
thinking about how models think • prev @MIT cs + physics catch me at https://t.co/YHXYaxexnU
Princeton, NJ
Joined October 2013
1/7 Wondered what happens when you permute the layers of a language model? In our recent paper with @tegmark, we swap and delete entire layers to understand how models perform inference - in doing so we see signs of four universal stages of inference!
22
102
566
RT @DrJimFan: This is the most gut-wrenching blog I've read, because it's so real and so close to heart. The author is no longer with us. I…
0
415
0
RT @METR_Evals: How close are current AI agents to automating AI R&D? Our new ML research engineering benchmark (RE-Bench) addresses this q…
0
179
0
RT @dbaek__: 1/6 New paper! “The Geometry of Concepts: Sparse Autoencoder Feature Structure.” We find that the concept universe of SAE feat…
0
186
0
RT @JoshAEngels: 1/11: New paper! "Decomposing the Dark Matter of Sparse Autoencoders." We find that SAE errors and error norms are linear…
0
38
0
@gershbrain Maybe the idea of using skip connections in networks? There is a bit written about this in the History section of
0
0
0
RT @tslwn: We centered residual stream activation vectors before computing cosine similarities to control for changes in the mean L2 norm,…
0
1
0
RT @katie_m_collins: [New preprint!] What does it take to build machines that **meet our expectations** and **compliment our limitations**?…
0
60
0
If you’re at #ICML2024 swing by the Mechanistic Interpretability Workshop to chat :)
1/7 Wondered what happens when you permute the layers of a language model? In our recent paper with @tegmark, we swap and delete entire layers to understand how models perform inference - in doing so we see signs of four universal stages of inference!
0
1
22
@CalcCon @MickeyShaughnes I think I did come across this in my literature review - correct me if I’m wrong but I believe your work is look at phases of training dynamics. My work is looking at phases between the layers of a fully trained LLM. This is really interesting work!!
2
0
0
@CalcCon @MickeyShaughnes you said "this particular group has chosen to ignore our work" - I was just suggesting that I'm happy to cite it if you message me the works you felt were ignored :)
0
0
1
@CalcCon @MickeyShaughnes hey @CalcCon just saw this - feel free to DM the paper(s) you feel relevant and I’ll be happy to add it to the arxiv. I tried really hard to do a good literature review but I don’t doubt that I missed things :)
1
0
1
@A_K_Nain hi aakash! love the work - nice to see that we came to similar conclusions in our experiments with other model families as well would love to hear your thoughts :)
1/7 Wondered what happens when you permute the layers of a language model? In our recent paper with @tegmark, we swap and delete entire layers to understand how models perform inference - in doing so we see signs of four universal stages of inference!
1
0
6
RT @tegmark: I’m excited about our recent AI paper. I was quite surprised by LLM’s robustness to deleting entire layers, but it makes sense…
0
13
0
1/7 Wondered what happens when you permute the layers of a language model? In our recent paper with @tegmark, we swap and delete entire layers to understand how models perform inference - in doing so we see signs of four universal stages of inference!
0
0
1
RT @s_scardapane: *The Remarkable Robustness of LLMs: Stages of Inference?* by @vedanglad @wesg52 @tegmark Interesting series of experime…
0
9
0
@QuintinPope5 @giffmana saw that - an interesting paper! we cite the result and compare it to what happens when you swap the same layer :)
0
0
1
@Butanium_ @tegmark hey thanks! I love the llamas work in english paper! - (pretty sure I cited it!) I'll give the newer paper a read too - I think the Llama family results would look similar (I've run subsets of the phases) but will try and get it running on llama 7b soon!
1
0
2