Jingcheng Hu, Yinmin Zhang, Shijie Shang, Xiaobo Yang, Yue Peng, Zhewei Huang, Hebin Zhou, Xin Wu, Jie Cheng, Fanqi Wan, Xiangwen Kong, Chengyuan Yao, Kaiwen Yan, Ailin Huang, Hongyu Zhou, Qi Han, Zheng Ge, Xiangyu Zhang, Heung-Yeung Shum. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Xiangfeng Wang, Hangyu Guo, Yanlin Lai, Mitt Huang, Liang Zhao, Chengyuan Yao, Yinmin Zhang, Qi Han, Xiaoxiaoren, Chun Yuan, Tong Xu, Zheng Ge, Xiangyu Zhang, Daxin Jiang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Recent research has increasingly documented that large language models (LLMs) can exhibit social bias in consequential social contexts such as education. Yet this literature often evaluates LLM bias in isolation without calibrating it against human bias, even though human bias has been extensively documented and constitutes one important source of LLM bias through pretraining on human-generated text. Such calibration is important for understanding when and how LLMs may reproduce human patterns of bias, with strong practical implications for deploying general-purpose LLMs and customizing models on task-specific datasets. The current study examines these issues in postsecondary education contexts, comparing racial and gender bias in human- and LLM-generated responses to postsecondary online discussion posts. Using a small pilot dataset sampled at a public four-year institution in the United States, we find that humans exhibit small and insignificant levels of social bias and that LLM bias does not differ significantly from this human baseline. These analyses motivate a larger-scale investigation that may inform debates about the appropriateness of AI in online learning environments and the risks of fine-tuning LLMs on local educational data.

Self-regulated learning (SRL) is critical for student success in online postsecondary education. Many technology-based interventions have been studied to improve SRL skills, but few were situated in broad-access institutions that disproportionately serve systemically marginalized student populations in STEM fields. This study presents preliminary findings from a rapid-cycle evaluation that tests two technology-supported instructional strategies (videos and prompts) designed to improve SRL in online learning. Using fine-grained clickstream data from 141 students across ten sections of five courses taught at a minority-serving community college, we generate measures of SRL behavior and correlate them with students' exposure to tested strategies. Our results indicate modestly positive relationships between both videos and prompts and SRL behavior. In addition, prompts are more strongly correlated with SRL behavior for first-generation and female students than for their peers. These initial findings reveal the promise and complexity of implementing effective and equitable technology-supported interventions to develop SRL skills and mindsets among diverse student populations in online STEM education.

Adversarial training (AT) methods are effective against adversarial attacks, yet they introduce severe disparity of accuracy and robustness between different classes, known as the robust fairness problem. Previously proposed Fair Robust Learning (FRL) adaptively reweights different classes to improve fairness. However, the performance of the better-performed classes decreases, leading to a strong performance drop. In this paper, we observed two unfair phenomena during adversarial training: different difficulties in generating adversarial examples from each class (source-class fairness) and disparate target class tendencies when generating adversarial examples (target-class fairness). From the observations, we propose Balance Adversarial Training (BAT) to address the robust fairness problem. Regarding source-class fairness, we adjust the attack strength and difficulties of each class to generate samples near the decision boundary for easier and fairer model learning; considering target-class fairness, by introducing a uniform distribution constraint, we encourage the adversarial example generation process for each class with a fair tendency. Extensive experiments conducted on multiple datasets (CIFAR-10, CIFAR-100, and ImageNette) demonstrate that our BAT can significantly outperform other baselines in mitigating the robust fairness problem (+5-10\% on the worst class accuracy)(Our codes can be found at https://github.com/silvercherry/Improving-Robust-Fairness-via-Balance-Adversarial-Training).

Reliable evaluation of adversarial defenses is a challenging task, currently limited to an expert who manually crafts attacks that exploit the defense’s inner workings, or to approaches based on ensemble of fixed attacks, none of which may be effective for the specific defense at hand. Our key observation is that adaptive attacks are composed from a set of reusable building blocks that can be formalized in a search space and used to automatically discover attacks for unknown defenses. We evaluated our approach on 24 adversarial defenses and show that it outperforms AutoAttack, the current state-of-the-art tool for reliable evaluation of adversarial defenses: our tool discovered significantly stronger attacks by producing 3.0%-50.8% additional adversarial examples for 10 models, while obtaining attacks with slightly stronger or similar strength for the remaining models.