wassname/steer-heal-love
1
0
Fork 0
mirror of https://github.com/wassname/steer-heal-love.git synced 2026-06-27 17:02:34 +08:00
Code Issues Packages Projects Releases Wiki Activity

Update README.md

Browse Source
This commit is contained in:
wassname (Michael J Clark)
2026-06-10 16:24:58 +08:00
committed by GitHub
parent 00d5e9e4c3
commit 5c44d3c1f3
1 changed files with 7 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+7 -5
View File
@@ -19,14 +19,16 @@ What if Lex Fridman is right?
>
> -- Lex Fridman, to Eliezer Yudkowsky 3 h 18 min into [Lex Fridman Podcast #368, "Dangers of AI and the End of Human Civilization"](https://podscript.ai/podcasts/lex-fridman-podcast/368-eliezer-yudkowsky-dangers-of-ai-and-the-end-of-human-civilization/) (03:18:03)
## What is this about:
## Authors note: What is this about:
- I show we can constrain coherence with a RMSE(KL).
- I show you can "heal" a steering iteration, using this, and thus apply steering multiple times
Briefly
This is interesting because it overcomes many of the limitations of steering. Steering vectors inject incoherence. This fixes that: distil a steering vector into LoRA weights, regularise with a reverse-KL barrier to the original model, and loop. 8 rounds, no coherence collapse.
- I show we can constrain adapter coherence with a RMSE(KL), and it works
- Further I show you can "heal" a steering iteration, using this, and thus apply steering multiple times
You can see the results and psudeocode at the end, but first let me show you Gemma's journel!
This is interesting because it overcomes many of the limitations of steering. Steering vectors inject incoherence. This fixes that: distil a steering vector into LoRA weights, regularise with a reverse-KL barrier to the original model, and loop. Below you can see 7 rounds, with no coherence collapse on the adapter.
Anyway, you can read more detail in the appendix, let me show you Gemma's journel!
## ⭐ gemma's diary ⭐