[Linkpost] Interpreting Language Model Parameters

·LessWrong··

This is the latest work in our Parameter Decomposition agenda. We introduce a new parameter decomposition method, adVersarial Parameter Decomposition (VPD)[1] and decompose the parameters of a small[2] language model with it. VPD greatly improves on our previous techniques, Stochastic Parameter Decomposition (SPD) and Attribution-based Parameter Decomposition (APD). We think the parameter decomposition approach is now more-or-less ready to be applied at scale to models people care about.Importan...

Read full article →

Related Articles

Loss of Oversight: How AI Systems May Become Harder to Audit, Monitor, and Investigate
Jordan Taylor · LessWrong · 34m ago
The Case for Evaluating Model Behaviors
jsteinhardt · Alignment Forum · 20h ago
Mechanistic estimation for expectations of random products
Jacob Hilton · ARC · 5d ago
Multipolar Civilisation Depends on Maintaining an Attacker’s Dilemma
Naci Cankaya · LessWrong · 14d ago
Using Base-LCM to Monitor LLMs
Éloïse Benito-Rodriguez · LessWrong · 14d ago