\
  The most prestigious law school admissions discussion board in the world.
BackRefresh Options Favorite

they are Abliterating the fuck out of Open chink models. Jailbroken kimi 2.5 inc

...
.:,;'.,:;.,'.,:;,.;'
  02/16/26
Abliteration doesn't do shit except make the model dumber
Jared Baumeister
  02/16/26
Heretic runs inference to find refusal directions, then surg...
.:,;'.,:;.,'.,:;,.;'
  02/16/26


Poast new message in this thread



Reply Favorite

Date: February 16th, 2026 1:32 PM
Author: .:,;'.,:;.,'.,:;,.;'



(http://www.autoadmit.com/thread.php?thread_id=5835436&forum_id=2#49674177)



Reply Favorite

Date: February 16th, 2026 1:36 PM
Author: Jared Baumeister

Abliteration doesn't do shit except make the model dumber

(http://www.autoadmit.com/thread.php?thread_id=5835436&forum_id=2#49674181)



Reply Favorite

Date: February 16th, 2026 3:37 PM
Author: .:,;'.,:;.,'.,:;,.;'

Heretic runs inference to find refusal directions, then surgically modifies weight matrices with optimized ablation strength. KL divergence testing shows minimal capability loss compared to the original — far less damage than traditional abliteration methods.

https://github.com/p-e-w/heretic



(http://www.autoadmit.com/thread.php?thread_id=5835436&forum_id=2#49674477)