If you look at elements of the residual stream that correspond to a specific token, and then see what MLP inputs map to what MLP outputs, how much of that can be described as a simple program that compares with token directions are present and then performs a shallow program to produce the output token directions?
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{holtzman-mapping-an-mlp-2026,
author = {Holtzman, Ari},
title = {Mapping an MLP in Terms of Tokens},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/jEMWwJOLbyG4DsqfcPns}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!