Papers by Xiaomeng Zhu
Learning Personalized Alignment for Evaluating Open-ended Text Generation (2024.emnlp-main)
Copied to clipboard
| Challenge: | Traditional evaluation metrics rely heavily on lexical similarity with human-written references, showing poor correlation with human judgments and failing to account for alignment with the diversity of human preferences. |
| Approach: | They propose an interpretable evaluation framework that evaluates alignment with specific human preferences by providing detailed comments and fine-grained scoring. |
| Outcome: | The proposed framework outperforms GPT-4 in Kendall correlation and accuracy with zero-shot reviewers. |
Parallelism and Generation Order in Masked Diffusion Language Models: Limits Today, Potential Tomorrow (2026.findings-acl)
Copied to clipboard
Yangyang Zhong, Yanmei Gu, Zhengqing Zang, Xiaomeng Li, Yuqi Ding, Xibei Jia, Yuting Shen, Zhenzhong Lan, Liwang Zhu, Weiping Liu, Junlin Zhou, Haisheng Liu, Zhong Xin Yu, Pengxin Luo, Donglian Qi, Yunfeng Yan, Junbo Zhao
| Challenge: | Autoregressive (AR) language models dominate modern natural language processing due to strong likelihood-based training objectives and reliable left-to-right decoding. |
| Approach: | They characterize MDLM behavior along two dimensions: parallelism strength and generation order . authors propose a Generate-then-Edit paradigm that mitigates dependency loss . |
| Outcome: | The proposed model improves on tasks that require "backward information" the Generate-then-Edit paradigm improves parallel decoding efficiency while reducing dependency loss. |
Meaning Beyond Truth Conditions: Evaluating Discourse Level Understanding via Anaphora Accessibility (2025.acl-long)
Copied to clipboard
| Challenge: | Existing assessments of understanding at the lexical and sentence levels are limited to lexica and sentence level, but few of them target whether LLMs accurately represent and update states of natural language discourse. |
| Approach: | They propose anaphora accessibility as a diagnostic for assessing discourse understanding . they use a dataset inspired by theoretical research in dynamic semantics to evaluate human and LLM performance. |
| Outcome: | The proposed dataset shows that humans and LLMs align on some tasks and diverge on others. |
LIEDER: Linguistically-Informed Evaluation for Discourse Entity Recognition (2024.acl-long)
Copied to clipboard
| Challenge: | Existing studies have shown that large language models have basic, if imperfect, DE recognition abilities, but they lack the basic knowledge of the fundamental semantic properties that govern the introduction and subsequent reference to DEs. |
| Approach: | They propose a dataset that examines large language models’ knowledge of existence, uniqueness, plurality, and novelty. |
| Outcome: | The proposed dataset shows that large language models exhibit sensitivity to all of these properties except novelty, which demonstrates that they have yet to reach human-level language understanding abilities. |
CYCLE-INSTRUCT: Fully Seed-Free Instruction Tuning via Dual Self-Training and Cycle Consistency (2025.emnlp-main)
Copied to clipboard
Zhanming Shen, Hao Chen, Yulei Tang, Shaolin Zhu, Wentao Ye, Xiaomeng Hu, Haobo Wang, Gang Chen, Junbo Zhao
| Challenge: | Existing methods for instruction tuning rely on expensive human-annotated seed data or powerful external teacher models. |
| Approach: | They propose a framework that achieves fully seed-free instruction tuning by employing a dual self-training loop where two models are bootstrapped solely from raw, unlabeled text. |
| Outcome: | The proposed framework outperforms seed-driven back-translation baselines and achieves comparable performance to strongly supervised methods. |