TL;DR: LLMs like GPT-4o, Claude, and Gemini show significant bias in recognizing scientists—favoring highly-cited researchers while underrecognizing women and researchers from Africa, Asia, and Latin America. Training data provenance contributes to these disparities. 

Yixuan Liu, Abel Elekes, Jianglin Lu, Rodrigo Dorantes-Gilardi, Albert-Laszlo Barabasi.

The 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP Findings), 2025.

TL;DR: An analysis of 3,000+ "critical letters" (papers published to critique others) finds that highly-cited, interdisciplinary, and novel papers are most likely to receive formal criticism. Surprisingly, receiving such criticism has no observable effect on a paper's citation trajectory or the authors' subsequent productivity.

Bingsheng Chen, Dakota Murray, Yixuan Liu, Albert-László Barabási

(Under review), 2025

TL;DR: This paper proposes a new graph-language model that uses the "scale-free" property of real-world networks as a structural prior, finding it can be effectively approximated by simple k-nearest neighbor graphs. This unified approach eliminates artificial assumptions about edge distributions and reduces the need for extensive labeled data.

Jianglin Lu, Yixuan Liu, Yitian Zhang, Yun Fu

The Thirteenth International Conference on Learning Representations (ICLR), 2025.

TL;DR: Using 90 years of IMDb data on 520K individuals, this study finds that winning film awards is strongly linked to career length, participating in previously awarded films, and collaborating with prize-winning actors. Interestingly, working with fewer directors (deeper relationships) increases award probability, while collaborating with more awarded actors also helps—suggesting Matthew effect in movies.

Yixuan Liu, Yifang Ma

Frontiers in Physics, 2022