If you ask an LLM to help you write something, especially a message, it often doesn't consider the pragmatic implications of what it suggests. Yet, LLMs are known to be able to read off these implications when asked. Can we probe for 'the variables under consideration' by an LLM by prefilling different responses that varies certain axes that might have been considered and sees if they change perplexity significantly?
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{holtzman-finding-variables-under-2026,
author = {Holtzman, Ari},
title = {Finding Variables Under Consideration},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/vxgsjgCnsmNL3S8eizBY}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!