Hal Daumé III’s research while affiliated with Loyola University Maryland and other places

... However, most VQA efforts are focused on evaluating machine "understanding," an abstract task which seeks to test the spatial or object reasoning of machine learning models. The resulting systems aren't easily extendable to an accessibility application [2]. The VizWiz dataset is an exception, where image-question pairs are obtained from the actual use case of blind and low-vision people asking questions about information in the visual world they're taking pictures of [7]. ...

Reference:

Context-VQA: Towards Context-Aware and Purposeful Visual Question Answering

... They encode observations so far as state, unused features/base classifiers as action space, and formulate various reward functions to account for classification error and costs. Such methods have been successfully applied in NLP He et al. (2013) to adaptively select features for dependency parsing, as well as computer vision Weiss and Taskar (2013) for human pose tracking. These methods use RL as a tool to learn decision rules that reduce test-time cost in the static environment of classification or structured prediction problems. ...

Reference:

Sequential Dynamic Decision Making with Deep Neural Nets on a Test-Time Budget

... These findings reinforce past work which argues that allocation harm is manifested in the additional effort is required by non-White users to adapt their prompts for successful use (Cunningham et al., 2024) and in the extra cost of additional tokens required to model linguistic variety (Ahia et al., 2023). While removing identity features from prompts altogether eliminates stereotypes from the recommendations, the sanitized outputs are still biased toward one group of users, providing less aligned services for those who do not identify with the model's default assumptions. ...

Reference:

Stereotype or Personalization? User Identity Biases Chatbot Recommendations

... This lack of ChatGPT Social AI influences users' moral judgments. This is true whether or not users know that the source of advice is generated by Social AI Si et al. (2023) Humans' ability to assess (true or false) information and explanations generated by Social AI compared to information snippets from Wikipedia N = 1500 participants across five conditions. The sample demographic is not disclosed ChatGPT Users tend to over-rely on explanations provided by ChatGPT, which decreases their ability to detect false claims in informational content Spitale et al. (2023) To what extent Social AI is more capable than humans at writing convincing, yet wrong information N = 697. ...

Reference:

Understanding model power in social AI

... GPT models represent available autoregressive decoder in-context learning via prompt engineering in 2023. Models have scaled and improved since then and it is possible that performance would improve, but issues of underlying racial biases (Section 5.5) continue to exist, even with more current models (Hofmann et al., 2024b,a;Warr et al., 2024;Shieh et al., 2024;Nghiem et al., 2024;Henderson et al., 2024). ...

Reference:

"All that Glitters": Approaches to Evaluations with Unreliable Model and Human Annotations

... Accurate review scores are crucial to improving peer review in very large-scale publication venues because they are the most important factor in determining accept/reject decisions. 1 The Isotonic Mechanism is well-suited for contemporary machine learning conferences, where it is commonplace for an author to submit multiple papers at the same time (Sun, 2020;Rastogi et al., 2022). An advantage of this methodology is that it requires minimal effort from authors and imposes no additional burden on reviewers. ...

Reference:

Analysis of the ICML 2023 Ranking Data: Can Authors' Opinions of Their Own Papers Assist Peer Review in Machine Learning?

... Here, we are interested in how the bias of a human-AI assemblage relates to the bias of humans-alone or AI-alone. Prior work on human-AI decision-making has found that the bias of a human-AI team is not simply equal to the sum of its parts and can depend on factors such as the decision-making task and whether or how the AI's suggestions are justified [e.g., 18,30,57,62,67]. Our paper considers the task of human text authorship with the help of word-level suggestions given by a predictive text system. ...

Reference:

Anti-stereotypical Predictive Text Suggestions Do Not Reliably Yield Anti-stereotypical Writing

... (2) We construct a realworld dataset, called HateDebias, where any sequence of sub-dataset in HateDebias corresponds to continuous-varying bias. (3) We propose a simple yet effective framework for continuous debiasing tasks, which provides a potential direction to improve continuous debiasing tasks. ...

Reference:

HateDebias: On the Diversity and Variability of Hate Speech Debiasing

... Large Language Models (LLMs) have shown impressive performance in reasoning and remembering extensive knowledge (Radford et al., 2019;Brown et al., 2020;Achiam et al., 2023;Anthropic, 2024;Zhao et al., 2023b;Touvron et al., 2023a,b;Yang et al., 2023;. However, when facing domain-specific questions, especially the e-commerce domain that has many long-tail entities and frequently updated information, LLMs often have issues such as hallucinations (Shi et al., 2023c;Wang et al., 2023;Zhao et al., 2023a). Retrieval-Augmented Generation * Corresponding authors. ...

Reference:

BSharedRAG: Backbone Shared Retrieval-Augmented Generation for the E-commerce Domain

... These systems typically return a single answer without any supporting context or evidence, making it challenging for users to understand how the answer was derived [40], [41]. This lack of transparency can reduce user trust and con dence in the system, especially in critical applications such as healthcare or legal domains [42]. Overall, while traditional QA systems have been valuable in certain contexts, their limitations have led to the development of more advanced approaches In recent years, the advent of powerful language models, such as the Generative Pre-trained Transformer (GPT), has revolutionized the eld of natural language processing (NLP) and opened up new possibilities for conversational agents [35], [43], [44], [45]. ...

Reference:

VaxBot-HPV: A GPT-based Chatbot for Answering HPV Vaccine-related Questions