Social biases can manifest in language agency. However, very limited research has investigated such biases in Large Language Model (LLM)-generated content. In addition, previous works often rely on string-matching techniques to identify agentic and communal words within texts, falling short of accurately classifying language agency. We introduce the **Language Agency Bias Evaluation (LABE)** benchmark, which comprehensively evaluates biases in LLMs by analyzing agency levels attributed to different demographic groups in model generations. LABE tests for gender, racial, and intersectional language agency biases in LLMs on 3 text generation tasks: biographies, professor reviews, and reference letters. Using LABE, we unveil language agency social biases in 3 recent LLMs: ChatGPT, Llama3, and Mistral. We observe that: (1) LLM generations tend to demonstrate greater gender bias than human-written texts; (2) Models demonstrate remarkably higher levels of intersectional bias than the other bias aspects. (3) Prompt-based mitigation is unstable and frequently leads to bias exacerbation. Based on our observations, we propose **Mitigation via Selective Rewrite (MSR)**, a novel bias mitigation strategy that leverages an agency classifier to identify and selectively revise parts of generated texts that demonstrate communal traits. Empirical results prove MSR to be more effective and reliable than prompt-based mitigation method, showing a promising research direction.

Recent large-scale T2I models like DALLE-3 have made progress in reducing gender stereotypes when generating single-person images. However, significant biases remain when generating images with more than one person. To systematically evaluate this, we propose the **Paired Stereotype Test (PST)** framework, which queries T2I models to depict two individuals assigned with male-stereotyped and female-stereotyped social identities, respectively (e.g. “a CEO” and “an Assistant”). This contrastive setting often triggers T2I models to generate gender-stereotyped images. Using PST, we evaluate two aspects of gender biases – the well-known **bias in gendered occupation** and a novel aspect: **bias in organizational power**. Experiments show that **over 74% images generated by DALLE-3 display gender-occupational biases**. Additionally, compared to single-person settings, DALLE-3 is more likely to perpetuate male-associated stereotypes under PST. We further propose **FairCritic**, a novel and interpretable framework that leverages an LLM-based critic model to i) detect bias in generated images, and ii) adaptively provide feedback to T2I models for improving fairness. FairCritic achieves near-perfect fairness on PST, overcoming the limitations of previous prompt-based intervention approaches.

Social commonsense reasoning naturally involves both the verbal and non-verbal cues of a social interaction. It is important for Large Vision-Language Models (VLMs) to leverage both textual and visual information in performing tasks like social understanding and reasoning. However, while current LLMs have shown good social reasoning capabilities in textual context, whether they can effectively incorporate visual information in social comprehension remains under-explored. To narrow the gap, we first construct and propose a benchmark: V-Social, featuring well-aligned text and visual content, tailored to assess visual social commonsense for multimodal foundation models. Through experimenting with V-Social, we find that even the most advanced VLM, GPT-4o, often falls short in social commonsense reasoning. This highlights the critical need to enhance the social grounding of VLMs. One major obstacle for improving this is the lack of high-quality data with good reasoning process. To overcome this obstacle, we introduce V-AlphaSocial, a novel method that generates high-quality chain-of-thought reasoning paths from unlabeled data. We design a visual reasoning reward model to improve VLM, and then iteratively refine both the VLM and the reward model. Our extensive analysis showcases how our method enhances social commonsense reasoning, proposing an effective approach that facilitates deeper exploration into field.

Prompt-based “diversity interventions” are commonly adopted to improve the diversity of Text-to-Image (T2I) models depicting individuals with various racial or gender traits. However, will this strategy result in nonfactual demographic distribution, especially when generating real historical figures? In this work, we propose **DemOgraphic FActualIty Representation (DoFaiR)**, a benchmark to systematically quantify the trade-off between using diversity interventions and preserving demographic factuality in T2I models. DoFaiR consists of 756 meticulously fact-checked test instances to reveal the factuality tax of various diversity prompts through an automated evidence-supported evaluation pipeline. Experiments on DoFaiR unveil that diversity-oriented instructions increase the number of different gender and racial groups in DALLE-3’s generations at the cost of historically inaccurate demographic distributions. To resolve this issue, we propose **Fact-Augmented Intervention** (FAI), which instructs a Large Language Model (LLM) to reflect on verbalized or retrieved factual information about gender and racial compositions of generation subjects in history, and incorporate it into the generation context of T2I models. By orienting model generations using the reflected historical truths, FAI significantly improves the demographic factuality under diversity interventions while preserving diversity.

Large vision-language models (LVLMs), while proficient in following instructions and responding to diverse questions, invariably generate detailed responses even when questions are ambiguous or unanswerable, leading to hallucinations and bias issues. Thus, it is essential for LVLMs to proactively engage with humans to ask for clarifications or additional information for better responses. In this study, we aim to shift LVLMs from passive answer providers to proactive engaged partners. We begin by establishing a three-tiered hierarchy for questions of invalid, ambiguous, and personalizable nature to measure the proactive engagement capabilities of LVLMs. Utilizing this hierarchy, we create PIE, (ProactIve Engagement Evaluation) through GPT-4o and human annotators, consisting of 853 questions across six distinct, fine-grained question types that are verified by human annotators and accompanied with well-defined metrics. Our evaluations on indicate poor performance of existing LVLMs, with the best-performing open-weights model only achieving an Aggregate Align Rate (AAR) of 0.28. In response, we introduce MACAROON, self-iMaginAtion for ContrAstive pReference OptimizatiON, which instructs LVLMs to autonomously generate contrastive response pairs for unlabeled questions given the task description and human-crafted criteria. Then, the self-imagined data is formatted for conditional reinforcement learning. Experimental results show MACAROON effectively improves LVLMs’ capabilities to be proactively engaged (0.84 AAR) while maintaining comparable performance on general tasks.

In English speaking assessment, pretrained large language models (LLMs) such as BERT can score constructed response items as accurately as human raters. Less research has investigated whether LLMs perpetuate or exacerbate biases, which would pose problems for the fairness and validity of the test. This study examines gender and native language (L1) biases in human and automated scores, using an off-the-shelf (OOS) BERT model. Analyses focus on a specific type of bias known as differential item functioning (DIF), which compares examinees of similar English language proficiency. Results show that there is a moderate amount of DIF, based on examinees’ L1 background in grade band 912. DIF is higher when scored by an OOS BERT model, indicating that BERT may exacerbate this bias; however, in practical terms, the degree to which BERT exacerbates DIF is very small. Additionally, there is more DIF for longer speaking items and for older examinees, but BERT does not exacerbate these patterns of DIF.

The recent LLMs like GPT-4 and PaLM-2 have made tremendous progress in solving fundamental math problems like GSM8K by achieving over 90% accuracy. However, their capabilities to solve more challenging math problems which require domain-specific knowledge (i.e. theorem) have yet to be investigated. In this paper, we introduce TheoremQA, the first theorem-driven question-answering dataset designed to evaluate AI models’ capabilities to apply theorems to solve challenging science problems. TheoremQA is curated by domain experts containing 800 high-quality questions covering 350 theorems from Math, Physics, EE&CS, and Finance. We evaluate a wide spectrum of 16 large language and code models with different prompting strategies like Chain-of-Thoughts and Program-of-Thoughts. We found that GPT-4’s capabilities to solve these problems are unparalleled, achieving an accuracy of 51% with Program-of-Thoughts Prompting. All the existing open-sourced models are below 15%, barely surpassing the random-guess baseline. Given the diversity and broad coverage of TheoremQA, we believe it can be used as a better benchmark to evaluate LLMs’ capabilities to solve challenging science problems.

Syntactically controlled paraphrase generation requires language models to generate paraphrases for sentences according to specific syntactic structures. Existing fine-tuning methods on this task is costly, as all parameters of the model need to be updated during the training process. Inspired by recent studies on parameter-efficient learning, we propose Parse-Instructed Prefix (PIP), a novel adaptation of prefix-tuning to tune large pre-trained language models on syntactically controlled paraphrase generation task in a low-data setting with significantly less training cost. We introduce two methods to instruct a model’s encoder prefix to capture syntax-related knowledge: direct initiation (PIP-Direct) and indirect optimization (PIP-Indirect). Comparing to traditional fine-tuning methods for this task, PIP is a compute-efficient alternative with 10 times less learnable parameters. Comparing to existing prefix-tuning methods, PIP excels at capturing syntax control information, achieving significantly higher performance at the same level of learnable parameter count.

Large Language Models (LLMs) have recently emerged as an effective tool to assist individuals in writing various types of content, including professional documents such as recommendation letters. Though bringing convenience, this application also introduces unprecedented fairness concerns. Model-generated reference letters might be directly used by users in professional scenarios. If underlying biases exist in these model-constructed letters, using them without scrutinization could lead to direct societal harms, such as sabotaging application success rates for female applicants. In light of this pressing issue, it is imminent and necessary to comprehensively study fairness issues and associated harms in this real-world use case. In this paper, we critically examine gender biases in LLM-generated reference letters. Drawing inspiration from social science findings, we design evaluation methods to manifest biases through 2 dimensions: (1) biases in language style and (2) biases in lexical content. We further investigate the extent of bias propagation by analyzing the hallucination bias of models, a term that we define to be bias exacerbation in model-hallucinated contents. Through benchmarking evaluation on 2 popular LLMs- ChatGPT and Alpaca, we reveal significant gender biases in LLM-generated recommendation letters. Our findings not only warn against using LLMs for this application without scrutinization, but also illuminate the importance of thoroughly studying hidden biases and harms in LLM-generated professional documents.

Recent advancements in Large Language Models empower them to follow freeform instructions, including imitating generic or specific demographic personas in conversations. We define generic personas to represent demographic groups, such as “an Asian person”, whereas specific personas may take the form of specific popular Asian names like “Yumi”. While the adoption of personas enriches user experiences by making dialogue systems more engaging and approachable, it also casts a shadow of potential risk by exacerbating social biases within model responses, thereby causing societal harm through interactions with users. In this paper, we systematically study “persona biases”, which we define to be the sensitivity of dialogue models’ harmful behaviors contingent upon the personas they adopt. We categorize persona biases into biases in harmful expression and harmful agreement, and establish a comprehensive evaluation framework to measure persona biases in five aspects: Offensiveness, Toxic Continuation, Regard, Stereotype Agreement, and Toxic Agreement. Additionally, we propose to investigate persona biases by experimenting with UNIVERSALPERSONA, a systematically constructed persona dataset encompassing various types of both generic and specific model personas. Through benchmarking on four different models- including Blender, ChatGPT, Alpaca, and Vicuna- our study uncovers significant persona biases in dialogue systems. Our findings also underscore the pressing need to revisit the use of personas in dialogue agents to ensure safe application.

Recent advances in natural language processing and transformer-based models have made it easier to implement accurate, automated English speech assessments. Yet, without careful examination, applications of these models may exacerbate social prejudices based on gender and race. This study addresses the need to examine potential biases of transformer-based models in the context of automated English speech assessment. For this purpose, we developed a BERT-based automated speech assessment system and investigated gender and racial bias of examinees’ automated scores. Gender and racial bias was measured by examining differential item functioning (DIF) using an item response theory framework. Preliminary results, which focused on a single verbal-response item, showed no statistically significant DIF based on gender or race for automated scores.

Existing studies have demonstrated that adversarial examples can be directly attributed to the presence of non-robust features, which are highly predictive, but can be easily manipulated by adversaries to fool NLP models. In this study, we explore the feasibility of capturing task-specific robust features, while eliminating the non-robust ones by using the information bottleneck theory. Through extensive experiments, we show that the models trained with our information bottleneck-based method are able to achieve a significant improvement in robust accuracy, exceeding performances of all the previously reported defense methods while suffering almost no performance drop in clean accuracy on SST-2, AGNEWS and IMDB datasets.