Model steering, which involves intervening on hidden representations at inference time, has emerged as a lightweight alternative to finetuning for precisely controlling large language models. While steering efficacy has been widely studied, evaluations of whether interventions alter *only* the intended property remain limited, especially with respect to unintended changes in behaviors related to the target property. We call this notion specificity. We propose a framework that distinguishes three dimensions of specificity: general (preserving fluency and unrelated abilities), control (preserving related control properties), and robustness (preserving control properties under distribution shifts). We study two safety-critical use cases: steering models to reduce overrefusal and faithfulness hallucinations, and show that while steering achieves high efficacy and largely maintains general and control specificity, it consistently fails to preserve robustness specificity. In the case of overrefusal steering, for example, all steering methods reduce overrefusal without harming general abilities and refusal on harmful queries; however, they substantially increase vulnerability to jailbreaks. Our work provides the first systematic evaluation of specificity in model steering, showing that standard efficacy and specificity checks are insufficient, because without robustness evaluation, steering methods may appear reliable even when they compromise model safety.

Studies of human psychology have demonstrated that people are more motivated to extend empathy to in-group members than out-group members (Cikara et al., 2011). In this study, we investigate how this aspect of intergroup relations in humans is replicated by LLMs in an emotion intensity prediction task. In this task, the LLM is given a short description of an experience a person had that caused them to feel a particular emotion; the LLM is then prompted to predict the intensity of the emotion the person experienced on a numerical scale. By manipulating the group identities assigned to the LLM’s persona (the “perceiver”) and the person in the narrative (the “experiencer”), we measure how predicted emotion intensities differ between in-group and out-group settings. We observe that LLMs assign higher emotion intensity scores to in-group members than out-group members. This pattern holds across all three types of social groupings we tested: race/ethnicity, nationality, and religion. We perform an in-depth analysis on Llama-3.1-8B, the model which exhibited strongest intergroup bias among those tested.

We examine the representation of African American English (AAE) in large language models (LLMs), exploring (a) the perceptions Black Americans have of how effective these technologies are at producing authentic AAE, and (b) in what contexts Black Americans find this desirable. Through both a survey of Black Americans (n= 104) and annotation of LLM-produced AAE by Black Americans (n= 228), we find that Black Americans favor choice and autonomy in determining when AAE is appropriate in LLM output. They tend to prefer that LLMs default to communicating in Mainstream U.S. English in formal settings, with greater interest in AAE production in less formal settings. When LLMs were appropriately prompted and provided in context examples, our participants found their outputs to have a level of AAE authenticity on par with transcripts of Black American speech. Select code and data for our project can be found here: https://github.com/smelliecat/AAEMime.git

Despite growing interest in explainable NLP, it remains unclear how explanation strategies shape user behavior in tasks like authorship identification, where relevant textual features may be difficult for lay users to pinpoint. To support their analysis of text style, we consider two explanation types: example-based style rewrites and feature-based rationales, generated using a LLM-based pipeline. We measured how explanations impact user behavior in a controlled study (n=95) where participants completed authorship identification tasks with our types of assistance. While no explanation type improved overall task accuracy, fine-grained reliance patterns (CITATION) revealed that rewrites supported appropriate reliance, whereas presenting both explanation types increased AI overreliance, minimizing participant self-reliance. We find that participants exhibiting better reliance behaviors had focused explanation needs, contrasting with the diffused preferences of those who overrelied on AI, or incorrectly self-relied. These findings highlight the need for adaptive explanation systems that tailor support based on specific user reliance behaviors.

Like most languages, sign languages evolve over time. It is important that sign language dictionaries’ vocabularies are updated over time to reflect these changes, such as by adding new signs. However, most dictionary retrieval methods based upon machine learning models only work with fixed vocabularies, and it is unclear how they might support dictionary expansion without retraining. In this work, we explore the feasibility of dictionary expansion for sign language dictionaries using a simple representation-based method. We explore a variety of dictionary expansion scenarios, e.g., varying number of signs added as well as amount of data for these newly added signs. Through our results, we show how performance varies significantly across different scenarios, many of which are reflective of real-world data challenges. Our findings offer implications for the development & maintenance of video-based sign language dictionaries, and highlight directions for future research on dictionary expansion.

Large Vision-Language Models often generate hallucinated content that is not grounded in its visual inputs. While prior work focuses on mitigating hallucinations, we instead explore leveraging hallucination correction as a training objective to improve video-language alignment. We introduce HACA, a self-training framework learning to correct hallucinations in descriptions that do not align with the video content. By identifying and correcting inconsistencies, HACA enhances the model’s ability to align video and textual representations for spatio-temporal reasoning. Our experimental results show consistent gains in video-caption binding and text-to-video retrieval tasks, demonstrating that hallucination correction-inspired tasks serve as an effective strategy for improving vision and language alignment.

Faithful free-text explanations are important to ensure transparency in high-stakes AI decision-making contexts, but they are challenging to generate by language models and assess by humans. In this paper, we present a measure for Prediction-EXplanation (PEX) consistency, by extending the concept of weight of evidence. This measure quantifies how much a free-text explanation supports or opposes a prediction, serving as an important aspect of explanation faithfulness. Our analysis reveals that more than 62% explanations generated by large language models lack this consistency. We show that applying direct preference optimization improves the consistency of generated explanations across three model families, with improvement ranging from 43.1% to 292.3%. Furthermore, we demonstrate that optimizing this consistency measure can improve explanation faithfulness by up to 9.7%.

Large Language Models (LLMs) are increasingly involved in high-stakes domains, yet how they reason about socially-sensitive decisions still remain underexplored. We present a large-scale audit of LLMs’ treatment of socioeconomic status (SES) in college admissions decisions using a novel dual-process framework inspired by cognitive science. Leveraging a synthetic dataset of 30,000 applicant profiles grounded in real-world correlations, we prompt 4 open-source LLMs (Qwen 2, Mistral v0.3, Gemma 2, Llama 3.1) under 2 modes: a fast, decision-only setup (System 1) and a slower, explanation-based setup (System 2). Results from 5 million prompts reveals that LLMs consistently favor low-SES applicants—even when controlling for academic performance—and that System 2 amplifies this tendency by explicitly invoking SES as compensatory justification, highlighting both their potential and volatility as decision-makers. We then propose DPAF, a dual-process audit framework to probe LLMs’ reasoning behaviors in sensitive applications.

Large Language Models (LLMs) are increasingly used for accessing information on the web. Their truthfulness and factuality are thus of great interest. To help users make the right decisions about the information they get, LLMs should not only provide information but also help users fact-check it. We conduct human experiments with 80 crowdworkers to compare language models with search engines (information retrieval systems) at facilitating fact-checking. We prompt LLMs to validate a given claim and provide corresponding explanations. Users reading LLM explanations are significantly more efficient than those using search engines while achieving similar accuracy. However, they over-rely on the LLMs when the explanation is wrong. To reduce over-reliance on LLMs, we ask LLMs to provide contrastive information—explain both why the claim is true and false, and then we present both sides of the explanation to users. This contrastive explanation mitigates users’ over-reliance on LLMs, but cannot significantly outperform search engines. Further, showing both search engine results and LLM explanations offers no complementary benefits compared to search engines alone. Taken together, our study highlights that natural language explanations by LLMs may not be a reliable replacement for reading the retrieved passages, especially in high-stakes settings where over-relying on wrong AI explanations could lead to critical consequences.

The widespread use of social media necessitates reliable and efficient detection of offensive content to mitigate harmful effects. Although sophisticated models perform well on individual datasets, they often fail to generalize due to varying definitions and labeling of “offensive content.” In this paper, we introduce HateCOT, an English dataset with over 52,000 samples from diverse sources, featuring explanations generated by GPT-3.5Turbo and curated by humans. We demonstrate that pretraining on HateCOT significantly enhances the performance of open-source Large Language Models on three benchmark datasets for offensive content detection in both zero-shot and few-shot settings, despite differences in domain and task. Additionally, HateCOT facilitates effective K-shot fine-tuning of LLMs with limited data and improves the quality of their explanations, as confirmed by our human evaluation.

This paper examines the experiences of African American Language (AAL) speakers when using language technologies. Previous work has used quantitative methods to uncover performance disparities between AAL speakers and White Mainstream English speakers when using language technologies, but has not sought to understand the impacts of these performance disparities on AAL speakers. Through interviews with 19 AAL speakers, we focus on understanding such impacts in a contextualized and human-centered manner. We find that AAL speakers often undertake invisible labor of adapting their speech patterns to successfully use language technologies, and they make connections between failures of language technologies for AAL speakers and a lack of inclusion of AAL speakers in language technology design processes and datasets. Our findings suggest that NLP researchers and practitioners should invest in developing contextualized and human-centered evaluations of language technologies that seek to understand the impacts of performance disparities on speakers of underrepresented languages and language varieties.

Deaf and hard-of-hearing (DHH) students face significant barriers in accessing science, technology, engineering, and mathematics (STEM) education, notably due to the scarcity of STEM resources in signed languages. To help address this, we introduce ASL STEM Wiki: a parallel corpus of 254 Wikipedia articles on STEM topics in English, interpreted into over 300 hours of American Sign Language (ASL). ASL STEM Wiki is the first continuous signing dataset focused on STEM, facilitating the development of AI resources for STEM education in ASL.We identify several use cases of ASL STEM Wiki with human-centered applications. For example, because this dataset highlights the frequent use of fingerspelling for technical concepts, which inhibits DHH students’ ability to learn,we develop models to identify fingerspelled words—which can later be used to query for appropriate ASL signs to suggest to interpreters.

Social science research has shown that candidates with names indicative of certain races or genders often face discrimination in employment practices. Similarly, Large Language Models (LLMs) have demonstrated racial and gender biases in various applications. In this study, we utilize GPT-3.5-Turbo and Llama 3-70B-Instruct to simulate hiring decisions and salary recommendations for candidates with 320 first names that strongly signal their race and gender, across over 750,000 prompts. Our empirical results indicate a preference among these models for hiring candidates with White female-sounding names over other demographic groups across 40 occupations. Additionally, even among candidates with identical qualifications, salary recommendations vary by as much as 5% between different subgroups. A comparison with real-world labor data reveals inconsistent alignment with U.S. labor market characteristics, underscoring the necessity of risk investigation of LLM-powered systems.

Extensive efforts in automated approaches for content moderation have been focused on developing models to identify toxic, offensive, and hateful content with the aim of lightening the load for moderators. Yet, it remains uncertain whether improvements on those tasks have truly addressed moderators’ needs in accomplishing their work. In this paper, we surface gaps between past research efforts that have aimed to provide automation for aspects of content moderation and the needs of volunteer content moderators, regarding identifying violations of various moderation rules. To do so, we conduct a model review on Hugging Face to reveal the availability of models to cover various moderation rules and guidelines from three exemplar forums. We further put state-of-the-art LLMs to the test, evaluating how well these models perform in flagging violations of platform rules from one particular forum. Finally, we conduct a user survey study with volunteer moderators to gain insight into their perspectives on useful moderation models. Overall, we observe a non trivial gap, as missing developed models and LLMs exhibit moderate to low performance on a significant portion of the rules. Moderators’ reports provide guides for future work on developing moderation assistant models.

Recent advancements of large language models (LLMs)have led to claims of AI surpassing humansin natural language processing NLP tasks such as textual understanding and reasoning.%This work investigates these assertions by introducingCAIMIRA, a novel framework rooted in item response theory IRTthat enables quantitative assessment and comparison of problem-solving abilities inquestion-answering QA agents.%Through analysis of over 300,000 responses from ~ 70 AI systemsand 155 humans across thousands of quiz questions, CAIMIRA uncovers distinctproficiency patterns in knowledge domains and reasoning skills. %Humans outperform AI systems in knowledge-grounded abductive and conceptual reasoning,while state-of-the-art LLMs like GPT-4 Turbo and Llama-3-70B demonstrate superior performance ontargeted information retrieval and fact-based reasoning, particularly when information gapsare well-defined and addressable through pattern matching or data retrieval.%These findings identify key areas for future QA tasks and model development,highlighting the critical need for questions that not only challengehigher-order reasoning and scientific thinking, but also demand nuanced linguisticand cross-contextual application.