(I am sure others have observed this, but I haven't seen it written out).
J-lens is great for quickly studying steering vectors

The output of tells you how logits are locally effected (on average) by steering by at layer . For instance, in Qwen-27B the 'excessive praise' steering direction has as its top logits

Rank Token J-lens score

1 magnificent +0.8083

2 exquisite +0.7857

3 beautifully +0.7711

4 dazzling +0.7511

5 elegant +0.7315

In general, you can often produce some useful hypotheses quickly on a given direction using the logit effect. For instance, the steering direction with these logit effects:
Rank Token J-lens score

1 functionalities +0.8992

2 impactful +0.8605

3 frameworks +0.8422

4 workflows +0.8421

5 leveraging +0.7969

Has a bunch of corporate words, and in fact it does push the model towards a more corporate style of interpreting the world.

This has the nice advantage that it is actually causal-- the direction pushed up here are the actual tokens that are moving up (locally, in expected value) when you steer in this direction.


Side note: In general, I would recommend applying where is the diagonal gain and is a centering matrix on logits. This is a better approximation to the norm than simply .