If we train an LLM like Talkie with purposeful time leakage—i.e., we correctly filter data that ngrams and record keeping would capture, but we know some documents past the cutoff date make it—then can we use such a model to test how assumed knowledge is encoded and how sensitive LLMs are to such assumed knowledge?
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{holtzman-purposeful-leakage-study-2026,
author = {Holtzman, Ari},
title = {Purposeful Leakage Study},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/8jXSFsDMb9SgxfFDz3j2}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!