/AI6h ago

Stanford NLP's Christopher Potts challenges rebuttal claiming sparse autoencoders outperform simple baselines for LLM steering

The dispute centers on evaluations within the AXBENCH benchmark.

--0--
Original posts
Comments
Original post
Aryaman Arora@aryaman2020#678inAI

has anyone ever written a diss track of your paper

6:49 PM · Jun 3, 2026 · 19.3K Views
Sentiment
Sentiment building, check back later.
Cluster Engagement
-
Views
-
Comments
-
Reposts
-
Bookmarks
Expand data
Posts from X
Most Activity
Most ActivityTimeline
VIEWS1K
Christopher Potts@ChrisGPotts

@aryaman2020 Is "can" as in "I can make a lay-up" or as in "I can make a shot from center court"?

Aryaman Arora@aryaman2020

has anyone ever written a diss track of your paper

5hViews 1KLikes 8Bookmarks 0
LIKES15RETWEETS2
Christopher Potts@ChrisGPotts

@aryaman2020 @lateinteraction In any case, no AI allowed. As Kendrick Lamar presciently said back in 2015, "I can dig rapping, but a rapper just prompting? What the f*ck happened?"

Christopher Potts@ChrisGPotts

@aryaman2020 @lateinteraction I think if Jørgensen and Hansen challenge us to a rap battle, we should absolutely accept. Or are we supposed to challenge them first? I am not sure of the etiquette.

5hViews 992Likes 15Bookmarks 0
REPLIES1
Christopher Potts@ChrisGPotts

@aryaman2020 @lateinteraction I think if Jørgensen and Hansen challenge us to a rap battle, we should absolutely accept. Or are we supposed to challenge them first? I am not sure of the etiquette.

Aryaman Arora@aryaman2020

@lateinteraction i will let @ChrisGPotts post the steering vector-themed parody of Eminem's "Without Me" but at least i can reply with this

5hViews 191Likes 4Bookmarks 0
Stanford NLP's Christopher Potts challenges rebuttal claiming sparse autoencoders outperform simple baselines for LLM steering · Digg