If you force an LLM to just randomly include a word or end its sentence with a certain word, but not necessarily use them in the sentence, just kind of include them randomly in a grammatical way, almost as an aside. How much will this affect what they say? Will they actually be able to factor it out? I doubt it.
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{holtzman-local-forced-leakage-2026,
author = {Holtzman, Ari},
title = {Local Forced Leakage},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/VTDg86d5x5mi4UFJ0lQB}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!