Motivation: Communication does not always happen on the same abstraction level for both humans and AIs. [neologism]
Research Question: Understand how implicit assumptions in a human's instructions might affect the model's instruction-following behavior. How does the model solve these underspecification problems, and how does it understand “implicature”? Can we detect it and improve our instructions? This might also related to reward design in reinforcement learning. How explicit the instruction reward should be?
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{bai-how-does-model-2026,
author = {Bai, Xiaoyan},
title = {How does model understand implicature.},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/lf9FJjuq1MTyTcpDL13d}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!