Alright so I've been messing around with LLMs for a few weeks now. SAE features are supposed to be interpretable, but when I wanted to directly attack an AI's own ontology, the whole thing kinda broke down.

Linear probes find directions that work, but I didn't know WHY they work. So I combined them.

I train probes on SAE feature activations instead of raw activations. Now the probe weights tell you exactly which features matter and how much.

TL;DR

Get your contrastive examples
Get your SAE features
Train a sparse linear probe with Elasticnet (most features become zero straight away)
The probe weights tell you which features matter and which ones don't
Build a steering vector by combining those SAE decoder directions weighted by the probe coefficients
If your probe is dirty, you can remove features, invert them, or learn what you need to change in your data

Code is here: https://github.com/ZuiderveldTimJ/HELP It works on my machine.

The Problem

Activation steering works. You get a steering vector that makes an AI output funny things, But you're left with a bunch of numbers.

SAE features were supposed to be interpretable, but when I started learning about all this a few weeks ago, I wasn't sure what I was supposed to do with them. There's like 4 google ones and if I want to steer for a complex concept like "being reunited with a friend who was lost at sea" or "AI self-knowledge," what do I do?

The Solution: Supervised Feature Selection

Train an ElasticNet probe on SAE feature activations. The sparsity constraint does the work, most features get zero weight.

For me, features are search spaces. The probe learns which features distinguish examples and what weight to give each feature.

What you get with all this:

A probe that's:

Transparent: You get an explicit formula. Not a bunch of numbers, it's a ranked list of interpretable features with weights you can inspect, judge and change.

Debuggable: See strange correlations immediately. Feature #11485 (fortune cookies) has a weight of -0.016. the probe thinks fortune cookies matter, Am I wrong? Is the feature wrong? I don't know but now i know where to look.

Editable: The probe is just a weighted sum. Don't like a feature? Delete it from the JSON. Want to flip the sign? Multiply by -1. Rebuild the steering vector in 2 seconds.

Technical Details

Method

Extract SAE feature activations for positive/negative examples (last tok

...

Posts

Wikitag Contributions

Comments

Posts

Wikitag Contributions

Comments