I have a theory: I think that inference-time info (e.g. text an LLM conditions on) is more likely to override parametric info if it sounds very similar to parametric info, but isn't clearly in opposition to it. I think this causes the model register that knowledge as if it was 'obviously true' but refer back to the in-context version rather than the parametric version.
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{holtzman-similarity-eclipsing-2026,
author = {Holtzman, Ari},
title = {Similarity Eclipsing},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/gfaM74Dbh56s5NDKymfq}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!