Interesting results, thanks for sharing! To clarify, what exactly are you doing after identifying a direction vector? Projecting and setting its coordinate to zero? Actively reversing it?
And how do these results compare to the dumb approach of just taking the gradient of the logit difference at that layer, and using that as your direction?
Some ad-hoc hypotheses for what might be going on:
Epistemic status: I ran some experiments with surprising results, and I don’t have a clear intuition of what it means for the way concepts are encoded in model activations. Help is welcome!
[EDIT] I made a mistake while analyzing the data, and it changes some of the early interpretations of the results. It doesn't change the fact that using a classifier is a poor way of finding the directions which matter most, but it lifts most of the mystery.
[EDIT 2] I found a satisfying explanation of the phenomenon in Haghighatkhah, 2022. Helpful pictures are included in the paper.
The mystery
If you want to find the “direction” corresponding to a concept in a Large Language Model, “natural” methods fail to find a crucial direction: you need to remove >5 dimensions for the model’s ability to use the concept to be affected. This seems to show that the concept is encoded in “a large number of dimensions”. But adversarial methods are able to find one direction which is quite significant (large effect when “removed”). How is that? What does that mean about how information is encoded by LLM?
More details
Let’s say you want to remove the concept of gender from a GPT-2’s activations between layers 23 and 24 (out of 48). (Both are >100 dimensional vectors). A natural thing to do is to have a bunch of sentences about guys and girls, train a linear classifier to predict if the sentence is about guys or girls, and use the direction the linear classifier gives you as the “direction corresponding to the concept of gender”. Then, you can project GPT-2’s activation (at inference time) on the plane orthogonal to the direction to have the model not take gender into account anymore.
This naive attempt fails: you can train a classifier to classify sentences as “guy-sentence” or “girl-sentence” based on the projected activations, and it will get 90% accuracy (down from 95%), and if you project GPT-2’s activation in the middle of the network, it will make only a tiny difference in the output of the network. This is a method called INLP (Ravfogel, 2020) (the original technique removes multiple directions to obtain a measurable effect).
But if you jointly train the projection and the linear classifier in an adversarial setup, you will get very different results: accuracy drops to ~50%, and GPT-2’s behavior changes much more, and sometimes outputs “she” when “he” is more appropriate. This is called RLACE (Ravfogel, 2022).
Why is that?
The best explanation I have found to explain this discrepancy is that INLP (= using a classifier) finds the direction with
the largest differencea clear separation, and it has almost no impact on model behavior. While RLACE (= using an adversarialy trained classifier and projection) finds a different direction (cosine similarity of 0.6), a direction where there isa clear separationa large difference, and it has a much greater impact on model behavior.[Edit (original images were wrong: titles were swapped)]
[Edit] Here is what it looks like in the plane defined by the directions found by INLP and RLACE: INLP finds a clear separation, and RLACE uses changes with large magnitude as well as very large features.
Is this something special about the distribution of neural activations?
No: if you use another “natural” distribution like GLOVE word encodings (which Ravfogel’s papers study), you will find that INLP fails to remove linearly available information, while RLACE is much better at it.
However, that’s not that surprising given the explanation above: the directions where there is a crisp distinction between two opposite concepts are not the one along which the magnitude of the changes is larger.
But just because information is linearly available doesn’t mean the network will be able to use it, especially if the direction you remove is one where the magnitude of the activations is the large: I’m surprised by how little INLP affects model behavior.
What does that mean about how concepts are encoded?
I have already seen people claiming that the large features seen in LLMs are probably the sign that models are encoding crisp concepts. I think that this experiment provides evidence of the opposite: it seems that projecting along directions where activations differ a lot is less impactful than projecting along directions where activations differ less, but in a crisper way. I would have expected the opposite to be true, and I’m open to explanations of why networks don’t seem to use directions along which activations maximally vary.(This part of the mystery was entirely due to a mistake I made while analyzing the results.)[Edit] I have already seen people claiming that the large features seen in LLMs are probably the sign that models are encoding crisp concepts. This experiment seems to support this conclusion: projecting down the directions where there is a crisp distinction make little to no difference, but projecting down the directions where there are large activations seems to affect model behavior a lot: the model seems to not rely on the crisp feature it builds, and instead relies on other features with large magnitude.
References and notebooks with the experiments
Appendix
Is it something specific to gender?
I tried exactly one more concept: political left vs political right, and I got the exact same results. See here: https://colab.research.google.com/drive/1wXQiiqU9Iy0lPCC92JvtK20kH5HKfWQr?usp=sharing
[Edit] Is it something specific to model trained with dropout?
Question asked by Neel Nada
No, similar results were found with GPT-Neo-125M (which I think is trained without dropout). Results can be found here https://colab.research.google.com/drive/1we8CrqMjgDUW2n5HIqlcoSeZxxODqXul?usp=sharing
Note: in the experiments I ran on GPT-Neo-125M, RLACE found a direction with extremely large direction, and impacted abilities of the model on all themes. I don't know if it's related to dropout or not.
[Edit] Can you have a bigger impact by swapping instead of projecting to zero?
Question asked by Neel Nada
Yes! After the prompt "Ezra is such a prick! Yesterday", the original model uses uses "he" to describe the subject on 3/3 completion, 2/3 with the model where data is "projected to median along the direction", and 0/3 with the model where data is "swapped along the direction" (with RLACE). It doesn't change the fact that INLP has a very small impact. And it doesn't
See here for more examples https://colab.research.google.com/drive/1IHL7HIWo6Ez0QHajMaSRoQZRyTg-CEOA?usp=sharing
[Edit] Here are some generations
For each distinct prompts, each model generates three completions with different seeds (each line is done with the same seed).