Uses WildVis
A Course Correction in Steerability Evaluation: Revealing Miscalibration and Side Effects in LLMs
AAAI 2026 · First arXiv submission
Uses keyword searches and manual review in WildVis to select realistic goals for text rewriting. The study tests whether models can change one property, such as reading difficulty or formality, while controlling other properties.
Data used
English-language examples retrieved through WildVis validate the study’s text-rewriting goals; the controlled experiments use separate source-text corpora.
Findings
WildVis examples support the choice of rewriting goals. The experiments find miscalibration and unintended changes to other text properties.
Appendix D.5: WildVis searches, English filtering, manual review, and conversation examples. Read the source.
Dates and versions
First arXiv version. ArXiv version history.
AAAI 2026 proceedings paper.
Cite this paper
Trenton Chang, Tobias Schnabel, Adith Swaminathan, Jenna Wiens. 2026. A Course Correction in Steerability Evaluation: Revealing Miscalibration and Side Effects in LLMs. Proceedings of the AAAI Conference on Artificial Intelligence.
Show BibTeX
@inproceedings{chang2026coursecorrection,
title = {{A Course Correction in Steerability Evaluation: Revealing Miscalibration and Side Effects in LLMs}},
author = {Chang, Trenton and Schnabel, Tobias and Swaminathan, Adith and Wiens, Jenna},
year = {2026},
booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
publisher = {AAAI Press},
volume = {40},
number = {44},
pages = {37259--37267},
doi = {10.1609/aaai.v40i44.41057},
url = {https://doi.org/10.1609/aaai.v40i44.41057}
}