ExpressivityBench: Can LLMs Communicate Implicitly?

Joshua Tint, Som Sagar, Aditya Taparia, Kelly Raines, Bimsara Pathiraja, Caleb Liu, Ransalu Senanayake

Findings of the Association for Computational Linguistics: EACL, 2026

picture_as_pdf PDFdescription arXiv
Overview figure for ExpressivityBench, showing the generator-to-grader channel used to measure how much implicit signal a language model transmits.

Summary

ExpressivityBench measures how well large language models convey information implicitly, modelling generation as an information-theoretic channel between a generator and a grader across nine tasks spanning emotion, tone, age, gender, and political slant. We find model strengths are uneven and strongly task-dependent, with affective content handled far better than sociolinguistic signals.

BibTeX

@inproceedings{tint2026expressivitybench,
    title={ExpressivityBench: Can LLMs Communicate Implicitly?},
    author={Tint, Joshua and Sagar, Som and Taparia, Aditya and Raines, Kelly and Pathiraja, Bimsara and Liu, Caleb and Senanayake, Ransalu},
    booktitle={Findings of the Association for Computational Linguistics: EACL 2026},
    pages={4500--4515},
    year={2026}
}