Models can easily understand text past word substitutions, e.g., if I tell them "in the following text, flamingo->apple". How many can a model bear without significant degradation? How does CoT alleviate this? At what rate does it help? Is it linear in the CoT length?
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{holtzman-word-substitution-load-2026,
author = {Holtzman, Ari},
title = {Word Substitution Load},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/y4YB7t9VOSz0Q9ujTc37}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!