We are delighted to share that six papers from the lab have been accepted to ACL 2026, including three Main Conference Oral presentations.
-
Representation of Backchannels: Learning useful representations for these small conversational cues is both necessary and effective. This project went through many twists and turns; many thanks to first author Dr. Yu Wang and Leyi Lao for seeing it through.
-
Think in Sentences: What if language models thought in sentences, much like humans do? From an initial spark of insight to a fully realized set of experiments, Zhichen Liu pursued this idea with remarkable persistence—and it has paid off.
-
Periodicity of Information: Information entropy exhibits periodicity, with LLMs showing stronger periodic patterns than humans. The motivation may sound a little science-fictional, while the method is grounded in straightforward signal processing. We hope this work brings a fresh perspective to longstanding questions in psycholinguistics.
-
Purging the Gray Zone: This work uses the geometry of representation spaces to reduce hallucinations in large language models. The idea is intuitive and effective. We thank Ph.D. student Hao An for his continued persistence in leading undergraduates Yibin Lou and Jiayi Guo through this exploration, bringing a clear signal out of subtle structure.
-
WaveDetect: A tool for distinguishing LLM-generated text from human text using wavelet features. It builds on our earlier spectral view of information entropy—and performs surprisingly well!
Congratulations to all our students!