A Tension Between Abstraction and Grounding

25

I've noticed that if I don't specifically ask LLMs to ground certain observation in a text or an image that I've sent to the LLM, then it tends to invoke extra information, often leading to hallucination. But that if I do ask for that grounding, then it tends to be specific and monocausal, only making observations that can relate to exactly one part of the input document or context. Is this a fundamental tension in LLMs?

llms grounding hallucination

Chat

If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:

@misc{holtzman-a-tension-between-2026,
  author = {Holtzman, Ari},
  title = {A Tension Between Abstraction and Grounding},
  year = {2026},
  url = {https://hypogenic.ai/ideahub/idea/ODkePWQU2Cj2i6khBrdS}
}

Comments (0)

Please sign in to comment on this idea.

No comments yet. Be the first to share your thoughts!