Show LLMs a text-message exchange. Have it propose different interpretations of what's being communicated/what each party is thinking. How many different possibilities can models propose for various levels of ambiguity? How many interpretations do they miss that humans propose in similar circumstances? How often do either humans or LLMs miss a reasonable interpretation?
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{holtzman-coverage-of-interpretations-2026,
author = {Holtzman, Ari},
title = {Coverage of Interpretations},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/TNWSMLdK6JWBGB3t6KHJ}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!