Damin Zhang

Damin Zhang, Julia Rayz, Romila Pradhan

Recently, language models have demonstrated strong performance on various natural language understanding tasks. Language models trained on large human-generated corpus encode not only a significant amount of human knowledge, but also the human stereotype. As more and more downstream tasks have integrated language models as part of the pipeline, it is necessary to understand the internal stereotypical representation in order to design the methods for mitigating the negative effects. In this paper, we use counterexamples to examine the internal stereotypical knowledge in pre-trained language models (PLMs) that can lead to stereotypical preference. We mainly focus on gender stereotypes, but the method can be extended to other types of stereotype. We evaluate 7 PLMs on 9 types of cloze-style prompt with different information and base knowledge. The results indicate that PLMs show a certain amount of robustness against unrelated information and preference of shallow linguistic cues, such as word position and syntactic structure, but a lack of interpreting information by meaning. Such findings shed light on how to interact with PLMs in a neutral approach for both finetuning and evaluation.

Damin Zhang, Julia Rayz

Understanding time is fundamental to human cognition, where temporal experience is often conceptualized through spatial metaphors grounded in sensory-motor experience. For example, "summer is approaching" parallels "We are approaching the summer". In such expressions, humans rely on a frame of reference (FoR) to interpret meaning relative to a particular viewpoint. Extending this concept to time, a temporal frame of reference (t-FoR) defines how temporal relations are perceived relative to an experiencer's moment of "now". While Large Language Models (LLMs) have shown remarkable advances in natural language understanding, their ability to interpret and reason about time remains limited. In this work, we introduce TUuD (Temporal Understanding under Deictic t-FoR), a framework that evaluates how LLMs interpret time-event and event-event relations when the reference point of "now" dynamically shifts along a timeline. Following recent work on temporal cognition \cite{li2025other}, LLMs are prompted to rate the similarity between the current moment and a target event from 0.00 (completely dissimilar) to 1.00 (highly similar), where similarity quantifies perceived temporal alignment between the two points. Our results show that four evaluated LLMs exhibit measurable adaptation to a deictic t-FoR, with similarity ratings peaking around the present and decreasing toward past and future events. The adaptation, however, weakens beyond near-term contexts, suggesting that while LLMs display partial human-like temporal cognition, their temporal reasoning remains sensitive to reference-frame shifts and temporal distance.

Damin Zhang, Yi Zhang, Geetanjali Bihani et al.

With the impressive performance in various downstream tasks, large language models (LLMs) have been widely integrated into production pipelines, like recruitment and recommendation systems. A known issue of models trained on natural language data is the presence of human biases, which can impact the fairness of the system. This paper investigates LLMs' behavior with respect to gender stereotypes, in the context of occupation decision making. Our framework is designed to investigate and quantify the presence of gender stereotypes in LLMs' behavior via multi-round question answering. Inspired by prior works, we construct a dataset by leveraging a standard occupation classification knowledge base released by authoritative agencies. We tested three LLMs (RoBERTa-large, GPT-3.5-turbo, and Llama2-70b-chat) and found that all models exhibit gender stereotypes analogous to human biases, but with different preferences. The distinct preferences of GPT-3.5-turbo and Llama2-70b-chat may imply the current alignment methods are insufficient for debiasing and could introduce new biases contradicting the traditional gender stereotypes.

Damin Zhang

As large language models (LLMs) have been used in many downstream tasks, the internal stereotypical representation may affect the fairness of the outputs. In this work, we introduce human knowledge into natural language interventions and study pre-trained language models' (LMs) behaviors within the context of gender bias. Inspired by CheckList behavioral testing, we present a checklist-style task that aims to probe and quantify LMs' unethical behaviors through question-answering (QA). We design three comparison studies to evaluate LMs from four aspects: consistency, biased tendency, model preference, and gender preference switch. We probe one transformer-based QA model trained on SQuAD-v2 dataset and one autoregressive large language model. Our results indicate that transformer-based QA model's biased tendency positively correlates with its consistency, whereas LLM shows the opposite relation. Our proposed task provides the first dataset that involves human knowledge for LLM bias evaluation.