Rendered at 14:19:19 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
krackers 16 hours ago [-]
>J-lens beats a plain logit lens on some architectures and does nothing on others, and it isn't about size
The paper talked about this, the jacobian matrix corrects for the shift in basis from initial to final layer compared to logit lens which assumes that the residual remains in the same basis across layers. Maybe the latter is in fact true for some models/architectures so the J-lens doesn't do anything extra?
mune2gu-chan 7 hours ago [-]
I love the focus on the pre-answer thinking phase. It feels way more practical than just trying to steer the output with a generic system prompt.
ada1981 2 hours ago [-]
Yes I am curious if we will start to see tools that expose this surface more directly or built into the models.
great_psy 15 hours ago [-]
At least link the one made by anthropic. Can’t say if this is better or worse since I’m on mobile.
ada1981 15 hours ago [-]
I actually just saw the paper not the one they made. Just googled and found it.
The paper talked about this, the jacobian matrix corrects for the shift in basis from initial to final layer compared to logit lens which assumes that the residual remains in the same basis across layers. Maybe the latter is in fact true for some models/architectures so the J-lens doesn't do anything extra?
There’s is likely better!
https://www.neuronpedia.org/qwen3.6-27b/jlens?shareId=cmr2kx...