You get a few free runs if you register with an edu email. I tried once, it was ok - I think a truly successful approach would require some prompt iteration, so multiple runs. Also of course the result will very much depend on the problem at hand.

Much remains to be done on this front, ideas are welcome!

6/end

Finally, we wanted to ensure that what we're showing is LLM patterns not TunedLens patterns - see Appendix for the control analyses we do! (+a comparison with LogitLens) 5/

We then apply this approach to 3 case studies - prediction by part-of-speech, multi-token fact recall, and fixed-response question answering. Check the paper & Akshat's thread for details!

(look at this model predicting positive/negative sentiment - such a clear pattern!)

4/

Most early layer predictions get overturned! It appears that these are statistical guesses, made when the model has not yet processed enough contextual information.

(flip rate = how often the model's final prediction differs from the current layer's prediction)

3/

In early layers, most frequent tokens (e.g., "the") dominate predictions, whereas infrequent tokens gradually become more predicted later on.

2/

Perhaps! But language contains a lot of information about the physical world, so in principle these are learnable from language alone

Omg adorable! Thanks! :)

Our young lab greatly appreciates open datasets and code provided by pioneers in the field! @alexanderhuth.bsky.social
@amandalebel.bsky.social
@rjantonello.bsky.social
@mtoneva.bsky.social
@samnastase.bsky.social
@jixingli.bsky.social
@mschrimpf.bsky.social
@evfedorenko.bsky.social
etc
3/end

We have started with naturalistic audio stories and fMRI data, but the setup can be extended to other types of input modalities and brain data; feel free to contribute
github.com/GT-LIT-Lab/l...

2/