Uses WildChat
User Feedback in Human-LLM Dialogues: A Lens to Understand Users But Noisy as a Learning Signal
EMNLP 2025 · First arXiv submission
Examines the frequency, meaning, and usefulness of implicit feedback in real conversations. It separates feedback polarity from what users actually request, and tests whether regenerating responses with that additional context improves model training.
Data used
WildChat and LMSYS-Chat-1M, including dense manual annotations on 34 WildChat conversations and separate 1,000-conversation samples for feedback-conditioned response experiments.
Findings
Feedback semantics help on MTBench but do not consistently improve the longer, more complex WildBench tasks; positive feedback alone can be a noisy training signal.
Sections 3–4: manual labels and user behavior. Sections 5–7: regeneration and training. Table 1: the 34-conversation WildChat annotation sample. Read the source.
Dates and versions
First arXiv version. ArXiv version history.
EMNLP 2025 proceedings paper.
Cite this paper
Yuhan Liu, Michael J. Q. Zhang, Eunsol Choi. 2025. User Feedback in Human-LLM Dialogues: A Lens to Understand Users But Noisy as a Learning Signal. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing.
Show BibTeX
@inproceedings{implicituserfeedback2025,
title = {{User Feedback in Human-LLM Dialogues: A Lens to Understand Users But Noisy as a Learning Signal}},
author = {Liu, Yuhan and Zhang, Michael J. Q. and Choi, Eunsol},
year = {2025},
booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing},
publisher = {Association for Computational Linguistics},
pages = {2666--2681},
doi = {10.18653/v1/2025.emnlp-main.133},
url = {https://aclanthology.org/2025.emnlp-main.133/}
}