Cognition and Societal Benefits

Cognition and Societal Benefits is sponsored by Microsoft Research. Supports academic research that advances science, technology, culture, and society, focusing on improving human interactions via sociotechnical research.

AFMR: Cognition and Societal Benefits - Microsoft Research Human-computer interaction Human language technologies Data platforms and analytics Programming languages & software engineering Security, privacy & cryptography Medical, health & genomics Technology for emerging markets Events & academic conferences Microsoft Research podcast Microsoft Research newsletter Mixed Reality & AI - Cambridge Mixed Reality & AI - Zurich Accelerating Foundation Models Research Cognition and Societal Benefits Academic research plays such an important role in advancing science, technology, culture, and society.

This grant program helps ensure this community has access to the latest and leading AI models.

Brad Smith, Vice Chair and President AFMR Goal: Improve human interactions via sociotechnical research which increases trust, human ingenuity, creativity, and productivity, and decreases the digital divide while reducing the risks of developing AI which does not benefit individuals and society The proposals mainly focus on significant advancements in the field of healthcare, education, and various social aspects.

They highlight the use of Large Language Models (LLMs) to enhance several aspects, such as improved teaching in online education platforms, generating personalized education for cybersecurity, and advancing health outcomes research.

There are also proposals focused on understanding the proficiency of LLMs in extracting and understanding clinical data, simulating student interactions in classrooms, and developing privacy-aware medical dialogue systems. Other studies investigate the utility and harms of LLMs for mental health support, their use in English as a foreign language (EFL) education, and their potential application within the legal field.

In healthcare, LLMs aim to not only assist doctors in patient-trial matching and radiology report summarization but also to provide patients with more understandable health data. Additionally, there are efforts to align LLMs with the diversity of global user preferences, and establish standardized protocols for using Generative Artificial Intelligence (GAI) in behavioral research, among others.

A Foundation Model-Simulated Virtual Classroom for PreK-12 STEM Education 📝 George Mason University : Ziyu Yao (PI) The proposal is focused on using Large Language Models (LLMs) to simulate student agents discussing STEM concepts in a virtual classroom. The platform is intended to aid STEM concept learning in PreK-12 education, facilitating teacher professional development and immersive peer learning.

The researchers aim to develop student agents with consistent stances in concept understanding, which would interact and debate with each other. A human teacher or student can also partake in the discussions, fostering deeper concept learning. Beneath the Surface: How Large Language Models Reflect Hidden Bias (opens in new tab) Can LLMs Simulate Personas with Reversed Performance?

A Benchmark for Counterfactual Instruction Following (opens in new tab) Efficient but Vulnerable: Benchmarking and Defending LLM Batch Prompting Attack (opens in new tab) MATHVC:AnLLM-Simulated Multi-Character Virtual Classroom for Mathematics Education (opens in new tab) A Versatile Platform for Enhanced Teaching, Computing, and Data-Driven Research University of Illinois Urbana-Champaign : Volodymyr Kindratenko (PI) The proposal is about developing a versatile platform that allows the creation of course-specific chatbots for teaching and research purposes.

The system uses the GPT-4 model via OpenAI API, and is capable executing codes, accessing databases, and assisting with computational research tasks. Accelerating Health/AI Research with Improved Clinical De-identification 📝 MIT : Marzyeh Ghassemi (PI) Investigate the bias of de-identification systems on names in clinical notes via a large-scale empirical analysis.

To achieve this, we created 16 name sets that vary along four demographic dimensions: gender, race, name popularity, and the decade of popularity. We insert these names into 100 manually curated clinical templates and evaluate the performance of nine public and private de-identification methods. We found that there are statistically significant performance gaps along a majority of the demographic dimensions in most methods.

In the Name of Fairness: Assessing the Bias in Clinical Record De-identification (opens in new tab) Advancing Culturally Congruent Cancer Communications with Foundation Models Morehouse School of Medicine : Muhammed Idris (PI) Breast and Cervical Cancers are among the most common cancers and leaders in cancer-related deaths among women worldwide.

Due to low screening rates among disadvantaged groups, including minorities, late-stage diagnoses and mortality rates were higher when compared to their counterparts.

The overarching goal of our project is to evaluate the capabilities of foundation models to help facilitate accessible and culturally congruent cancer-related health information with the aim of addressing community concerns (i.e., trust, myths, access) and promoting cancer screenings among underrepresented women of color.

Specifically, we will evaluate and compare the breadth, quality, and accuracy of the health-related information around specific community concerns generated by GPT and Llama-2 model families and develop a prototype of an interactive tool, fine-tuned around specific community concerns related to breast and cervical cancer screening using GPT-3 and leverage Azure API and DALL-E 2 to address health literacy barriers.

AI-based Digital Twins of Doctors in Patient Care University of Illinois Chicago : Mohan Zalake (PI) The research aims to evaluate the acceptability of using Digital Twins of Doctors (DTDs) in patient care. DTDs are AI-generated characters that share the facial and vocal identities of real doctors and can deliver health information to patients.

Past research has explored the benefits (e.g., efficient delivery of repetitive information and personalizing patient care) and concerns (e.g., ethical and social concerns) for integrating DTDs in patient care from the perspective of doctors who share their identities with DTDs.

Given there exist both potential benefits and limitations, research efforts are required to systematically study the implications of using DTDs in healthcare with all the stakeholders before widely adopting them.

In this proposal, I aim to evaluate the acceptability of DTDs with the next important stakeholder: patients; by understanding a) how patients perceive and respond to DTDs that share the identities of their own doctors and b) how DTDs influence patients’ trust, engagement, comprehension, and adherence to health information and advice.

I will conduct a mixed-methods study involving patients interacting with DTDs, filling surveys, and follow-up interviews. The proposed research will contribute to the understanding of responsible integration of generative AI solutions like DTDs into healthcare.

Assessing Clinical Reasoning Capabilities of Large Language Models Through Simulated Doctor-Patient Conversations 📝 Harvard University : Pranav Rajpurkar (PI) This proposal aims to develop an evaluation framework for assessing the performance of Large Language Models in medical AI applications.

The framework simulates real-world doctor-patient conversations and uses AI agents to imitate doctor-patient interaction and assess conversational abilities. The project will initially focus on assessing the diagnosis of skin conditions.

An evaluation framework for clinical use of large language models in patient interaction tasks (opens in new tab) Direct Preference Optimization for Suppressing Hallucinated Prior Exams in Radiology Report Generation (opens in new tab) ReXPref-Prior: A MIMIC-CXR Preference Dataset for Reducing Hallucinated Prior Exams in Radiology Report Generation (opens in new tab) Assessing the potential and risks for accessibility and ableism in intelligent agents 📝 University of Washington : Jennifer Mankoff (PI) The proposal aims to explore the capability of Generative AI (GAI) in the realm of text simplification, particularly to aid individuals with cognitive impairments.

The team plans to develop a proof-of-concept text simplification system using Open AI’s GPT4 for both generation and validation, addressing the potential risks and consequences for people with disabilities along the way. The research seeks to bridge the digital divide for those who benefit from simplified text and in the process generate important new insights about the value of GAI in accessibility.

Autoethnographic Insights from Neurodivergent GAI “Power Users” (opens in new tab) Identifying and Improving Disability Bias in GPT-Based Resume Screening (opens in new tab) Causal Inference to Understand the Impact of Humanitarian Interventions on Food Security in Africa 📝 Universitat de Valencia : Gustau Camps-Valls (PI) The Causal4Africa project will investigate the problem of food security in Africa from a novel causal inference standpoint.

The project will illustrate the usefulness of causal discovery and estimation of effects from observational data by intervention analysis. Ambitiously, it will improve the usefulness of causal ML approaches for climate risk assessment by enabling the interpretation and evaluation of the likelihood and potential consequences of specific interventions.

Causal inference to study food insecurity in Africa (opens in new tab) Climate risk assessment needs urgent improvement (opens in new tab) Evaluating the Impact of Humanitarian Aid on Food Security (opens in new tab) Inferring causation from time series in Earth system sciences (opens in new tab) Large language models for causal hypothesis generation in science (opens in new tab) Small is beautiful: climate-change science as if people mattered (opens in new tab) Storylines for decision-making: climate and food security in Namibia (opens in new tab) Quantifying Causal Pathways of Teleconnections (opens in new tab) CausaLLM: Foundation Models for Real-World Evidence Generation 📝 University of California, Berkeley : Ahmed Alaa (PI) The proposal seeks to develop a foundation model that can aid in analyzing real-world observational data (RWD) and generate high-quality Real-World Evidence (RWE).

The researchers aim to reduce the development time and expert input required to devise Statistical Analysis Plans (SAPs). The steps proposed include studying the zero-shot performance of LLMs at generating SAPs, and then refining the process to produce a model that can automate SAP production.

Large Language Models as Co-Pilots for Causal Inference in Medical Studies (opens in new tab) Challenges and Benefits of AI Errors for Learners in Conversational Q&A Systems University of Illinois Urbana-Champaign : Karrie Karahalios (PI) The proposal aims to study the impact of AI errors on learners’ engagement, learning outcomes, and perceived helpfulness of AI educational systems.

The researchers intend to conduct an online controlled experiment involving adult learners in STEM where they are presented with various scenarios of imperfection in conversational Q&A systems. The outcomes of this study aim to provide the necessary knowledge to maximize learners’ gains and do so fairly.

Clinical Question and Answering Multimodal with Community Wound Infection Data University of British Columbia : Xiaoxiao Li (PI) This project aims to leverage Large Language Models (LLM) for multimodal medical data analysis in community healthcare focusing on wound care. It proposes to build a trustworthy conversational AI tuned to provide evidence-based responses to clinical inquiries.

Moreover, it seeks to overcome challenges related to multimodal data complexity, trustworthiness and fairness.

Copilot for Worker Wellbeing* 📝 Northeastern University : Vedant Swain (PI) * AICE Accelerator collaboration To make AI agents more empathetic towards worker’s goals, the agent needs to (i) understand broader wellbeing goals beyond saving time, (ii) maintain latitudinal and longitudinal awareness of workers’ context outside their task, and (iii) provide workers suggestions to meet those goals by preempting opportunities in their work context.

In this project, we propose to prototype and study Pro-Pilot, an enhancement over the existing Copilot that introduces a new Human-AI interaction framework that builds empathy.

AI on My Shoulder: Supporting Emotional Labor in Front-Office Roles with an LLM-based Empathetic Coworker (opens in new tab) SeSaMe: A Framework to Simulate Self-Reported Ground Truth for Mental Health Sensing Studies (opens in new tab) Teacher, Trainer, Counsel, Spy: How Generative AI can Bridge or Widen the Gaps in Worker-Centric Digital Phenotyping of Wellbeing (opens in new tab) Deidentification of Medical Data with Foundation Models​ University of Toronto : Alistair Johnson (PI) Create a highly accurate and efficient deidentification system that can be applied to various medical data sources, ultimately facilitating secure data sharing and collaboration in the healthcare industry.

Developing Foundation Models for Survival Prediction from Pathological Image and Biomedical Text The University of Texas at Arlington : Junzhou Huang (PI) This project proposes to address the critical challenge in personalized healthcare of accurately predicting survival outcomes using digital pathology techniques.

It identifies two key challenges: the complexity of microenvironment of tissues in histopathological images, and the integration of the images with corresponding biomedical text data. To tackle these, we propose two aims: The first is to develop an advanced cell segmentation foundation model that enhances feature extraction and analysis in histopathological images.

The second aim focuses on developing a multimodal foundation model that effectively combines pathological image features with biomedical captions for improved survival predictions.

These proposed methods promise to significantly influence both machine learning and histopathological imaging by introducing novel foundation models for integrated image-caption data analytics and have the potential to impact other related fields in a similar capacity.

Developing Privacy-Aware Medical Dialogue System Using Retrieval-Augmented Large Language Models (LLMs)​ 📝 Georgia Institute of Technology : May Dongmei Wang (PI) Expedite AI research and improve healthcare by developing a privacy-aware medical dialogue system that 1) leverages human interaction and prompting in dialogue systems for unstructured clinical data analysis, and 2) adapts large language models (LLMs) to clinical use cases.

EHRAgent: Code Empowers Large Language Models for Few-shot Complex Tabular Reasoning on Electronic Health Records (opens in new tab) Retrieval-Augmented Large Language Models for Adolescent Idiopathic Scoliosis Patients in Shared Decision-Making (opens in new tab) Dialog-based navigation model for visually impaired people​ 📝 Waseda University : Daisuke Kawahara (PI) Proposal for a system to assist visually impaired individuals in outdoor navigation by using vision and language foundation models to extract visual information from images/videos captured by mobile cameras.

These insights are communicated through relevant dialogues. Includes use of Azure and OpenAI technologies, and the creation of two specific datasets for model development.

CityNav: Language-Goal Aerial Navigation Dataset with Geographic Information (opens in new tab) JDocQA: Japanese Document Question Answering Dataset for Generative Language Models (opens in new tab) Map-based Modular Approach for Zero-shot Embodied Question Answering (opens in new tab) SlideAVSR: A Dataset of Paper Explanation Videos for Audio-Visual Speech Recognition (opens in new tab) Text360Nav: 360-Degree Image Captioning Dataset for Urban Pedestrians Navigation (opens in new tab) Vision Language Model-based Caption Evaluation Method Leveraging Visual Context Extraction (opens in new tab) Distributional Alignment of Large Language Models with the Diversity of Human Preferences University of Michigan, Ann Arbor : Qiaozhu Mei (PI) The proposal focuses on a novel approach, distributional alignment, that aligns Large Language Models (LLMs) to the broad spectrum of human preferences stemming from varied contexts.

This involves the collection of diverse human contexts and preferences across expansive domains and their integration into LLM training and refinement. Unique metrics to assess the diversity of LLM outputs will be introduced. The investigators request access to GPT-3, GPT-4, and other modeling resources to aid their research.

Electronic Health Record Summarization Harvard University : Emily Alsentzer (PI) Clinicians face challenges in summarizing a patient’s medical history upon hospital admission due to information overload in electronic health records.

We aim to develop LLM-based methods for generating factual summaries by leveraging retrieval-based approaches and to design evaluation approaches for assessing the quality of the generated summaries, comparing them to existing metrics and clinician evaluations.

EPIWATCH: an AI-driven system for early warnings of epidemics worldwide University of New South Wales : Raina MacIntyre (PI) The project aims to further develop EPIWATCH, an epidemic detection and surveillance system, using large language models (LLMs). The developed LLMs will be used to automoate certain functions of EPIWATCH and will include low-resource languages important for Australian communities.

The models will be fine-tuned and retrained for tasks including classification of public health threats and extraction of key information from large data sets. The project aims to fill the gap in AI usage for public health, especially for underrepresented languages.

Evaluating GPT’s Contract Generation Capabilities KAIST : Sangchul Park (PI) Study prompting or fine-tuning strategies for improving GPT’s capabilities for contract generation. A set of prompts can be produced as an output of this research. I find GPT particularly good at generating contract terms if proper prompts are fed into it.

I will study prompting or fine-tuning strategies for improving GPT’s capabilities for contract generation. A set of prompts can be produced as an output of this research. I will also try to prepare an evaluation set for measuring the performance of contract generation and compare GPT with other language models.

If there are difficulties in designing an evaluation set, I can consider, as an alternative, conducting a “snowball sampling” to solicit multiple law professors for human evaluation, or using my law school class to engage law students for evaluation.

Evaluating the use of GPT-4 to extract symptoms and medication use data from clinical notes 📝 University of California, San Francisco : Vivek Rudrapatna The project aims to evaluate the accuracy of GPT-4 in extracting patient symptoms and medications from clinical notes in the electronic health record.

It intends to compare GPT-4’s performance against comparator models and explore hybrid approaches to improve the accuracy of clinical information extraction.

Large language models outperform traditional natural language processing methods in extracting patient-reported outcomes in IBD (opens in new tab) Federated Privacy-Preserving Multimodal Generator for Synthetic Medical Data Generation San Diego State University : Hajar Homayouni AI’s potential in healthcare is limited by the scarcity of representative and balanced Electronic Health Records (EHRs).

This proposal aims to address issues stemming from inaccessible, incomplete, and biased EHRs crucial for critical data analysis and decision-making. The research approach involves utilizing limited available data to generate balanced, correlated EHRs for precise and equitable training and validation of data-driven models.

The proposed solution is a Federated Privacy-preserving Multimodal Generative (FPMG) framework, designed to generate unbiased EHR data and facilitate secure collaborative learning. Primarily, it targets the generation of balanced and correlated multimodal EHR data types, utilizing a deep generative adversarial model. By capturing cross-modal correlations and associations, the framework aims to enhance decision-making systems.

Additionally, the project seeks to explore decentralized cross-silo federated learning to safeguard patients data privacy and enhance the robustness and generalization of models in healthcare applications.

Foundation Models for Cultural Analytics and Computational Social Science 📝 University of California, Berkeley : David Bamman (PI) Foundation models such as ChatGPT, GPT-4 and Llama 2 are poised to transform research at the intersection of natural language processing and computational social science/cultural analytics, not simply in providing more accurate measuring instruments for existing tasks (Ziems et al.

2023) but also in opening up the ability to ask fundamentally more difficult questions that require world knowledge, long document context, and sophisticated inference.

This research project probes the ability of foundation models to accelerate responsible computational research in the social sciences and humanities; the goal is to generate new knowledge about culture and society and provide a roadmap for other researchers to do so themselves.

On Classification with Large Language Models in Cultural Analytics (opens in new tab) Foundation Models for Precise and Reliable Clinical Decision Making: A Study on Patient-trial Matching Rice University : Xia Hu (PI) The proposal seeks to leverage the abilities of Large Language Models (LLMs) to address the challenge of accurately and reliably matching patients with appropriate clinical trials.

It aims to achieve precise and reliable patient-trial matching by resolving the incompatibility between Electronic Health Records (EHRs) and clinical trial descriptions and providing comprehensive explanations for the matches.

Foundation Models for Socratic Conversations with Novice Debuggers 📝 University of North Carolina at Charlotte : Razvan Bunescu (PI) With the aim of improving teaching and learning of coding, we propose to develop Socratic conversational agents by fine-tuning large foundation models on a dataset of dialogues where an instructor helps students debug code.

The Socratic conversational agents are intended to augment human instruction, assisting novice programmers to fix their code and thus enhancing their learning outcomes. Can Language Models Employ the Socratic Method?

Experiments with Code Debugging (opens in new tab) Short Story Generation through Autoregressive Transfer of Narrative Continuations (opens in new tab) Foundation Models in Behavioral Science: A Guide to Ethical and Effective Use University of California, Berkeley : Juliana Schroeder (PI) This proposal tackles the urgent need for standardized protocols when integrating Generative Artificial Intelligence (GAI) into behavioral research.

The research goals include understanding the current state of GAI use in behavioral science, exploring the potential benefits and risks, and developing guidelines for its effective and responsible use. The proposal includes the conduction of a large field study and consultation with a panel of experts.

Foundational Models for Infrastructure Resilience: Use Case of Conversational AI for Disaster Recovery Florida International University : Mohammadhadi Amini (PI) Natural disasters introduce major challenges for critical infrastructure and human lives. In such scenarios, effective communication among first responders, agencies, and residents, is critical to ensure timely recovery and survival.

However, existing notification systems are not benefiting from the state-of-the-art AI-based solutions to handle the real-time evolving situations that arise during disasters. Hence, this project proposes to develop and evaluate a conversational agent, using Microsoft Azure OpenAI Services, that can facilitate the coordination of stakeholders in disaster situations using natural language.

The conversational agents that are created using Microsoft Azure OpenAI service will be based on large language models (LLMs) that are fine-tuned using historical disaster datasets. The PI and his team have prior experience in using pre-trained models for computer vision, critical infrastructure resilience, and healthcare applications. They used MS Azure services including Azure Machine Learning.

The performance of the conversational agent will be validated using synthetic scenarios that simulate realistic and challenging use cases and interactions in disasters, by tracking performance metrics such as loss, accuracy, and perplexity. The project will contribute to identifying and exploring new use cases for conversational AI in critical infrastructure resilience, and the outcomes will be disseminated via research publications.

Generative AI models for pandemic risk assessment Northeastern University : Samuel Scarpino (PI) This proposal sets out the goal of developing AI tools for pandemic risk assessment. This is to be achieved through a partnership between the CAPTRS and the IEAI at Northeastern University.

By using data from ProMED and the WHO Disease Outbreak Network, the team aims to fine-tune two distinct models – OpenAI’s GPT-3 text-davinci-003 and the open-source model Llama-2. The final outcome of this project would be models capable of generating early-stage outbreak alerts along with an associated risk score.

Hallucination Detection in Medical Text Generation: Enhancing Patient Safety and Care 📝 University of Cambridge : Mihaela van Der Schaar (PI) Develop and evaluate a hallucination detection system for medical text generation. By detecting and mitigating hallucinations in AI generated text, we aim to enhance patient safety and improve the quality of medical care.

Developing a hallucination detection system can improve the safety and quality of AI generated medical text. Furthermore, the insights gained from this research will contribute to the broader understanding of responsible AI deployment in healthcare and help develop best practices for the ethical use of AI in medicine.

Interpretable Medical Diagnostics with Structured Data Extraction by Large Language Models (opens in new tab) HealthGPT: Teaching and Utilizing GPTs for Trustworthy Health Outcomes Research 📝 Emory University : Carl Yang (PI) The proposed research aims to explore the application of foundation models, specifically GPTs, to advance health outcomes research.

This is achieved by focusing on their alignment with patient values, dealing with social determinants of health, and their capabilities in advancing health sciences and healthcare.

EHRAgent: Code Empowers Large Language Models for Few-shot Complex Tabular Reasoning on Electronic Health Records (opens in new tab) Knowledge-Infused Prompting: Assessing and Advancing Clinical Text Data Generation with Large Language Models (opens in new tab) PromptLink: Leveraging Large Language Models for Cross-Source Biomedical Concept Linking (opens in new tab) RAM-EHR: Retrieval Augmentation Meets Clinical Predictions on Electronic Health Records (opens in new tab) Uncertainty-Aware Pre-Trained Foundation Models for Patient Risk Prediction via Gaussian Process (opens in new tab) Implementing AI on an internal medicine ward for quick access to national guidelines University of Berlin : Matthias Groeschel (PI) The research project aims at evaluating the feasibility of GPT-4’s ability to compare physicians’ diagnostic and therapeutic choices to national guidelines.

The goal is to provide in hospital access to local and national guidelines for two chronic pulmonary diseases, and assessing and defining the prompt strategy for comparison between national guidelines and patient treatment. Improving Online Teaching: Applying GPT to Content from a Large K-12 Education Platform 📝 Carnegie Mellon University : Tom M.

Mitchell (PI) We propose research applying GPT models to improve the quality of teaching in online education platforms, and request Azure access to GPT to support this research. As a case study of this problem, we will partner with the widely used K-12 education platform freely available at the non-profit www. ck12.

org website, which has served over 100 million unique student visitors worldwide.

Ruffle&Riley: Towards the Automated Induction of Conversational Tutoring Systems (opens in new tab) Learning to Compare Hints: Combining Insights from Student Logs and Large Language Models (opens in new tab) In Plain English: Making Online Contracts Accessible to All University of Illinois Urbana-Champaign : Hari Sundaram (PI) Integrating ChatGPT into EFL Writing Education 📝 The integration of ChatGPT in the field of education has garnered significant interest, offering an opportunity to examine its effectiveness in English as a foreign language (EFL) education.

A novel learning platform, RECIPE, collects students’ interaction data with ChatGPT by guiding students and ChatGPT prompting. The goal is to investigate students’ usage and perception of generative AI and explore how students can effectively utilize generative AI in EFL writing education.

BLEND: A Benchmark for LLMs on Everyday Knowledge in Diverse Cultures and Languages (opens in new tab) Exploring Cross-Cultural Differences in English Hate Speech Annotations: From Dataset Construction to Analysis (opens in new tab) Exploring Student-ChatGPT Dialogue in EFL Writing Education (opens in new tab) LLM-C3MOD: A Human-LLM Collaborative System for Cross-Cultural Hate Speech Moderation (opens in new tab) RECIPE4U: Student-ChatGPT Interaction Dataset in EFL Writing Education (opens in new tab) Investigating the Utility and Challenges of Large Language Models (LLMs) for Improving Employee's Experience San Diego State University : Hossein Shirazi (PI) In the U.S. knowledge-based sector, 10% of new hires (around 2 million employees) face recurrent challenges, partly from a flawed mentorship system where less than 40% of employees receive mentorship.

Overwhelmed mentors hinder productivity, resulting in inefficiencies and wasted resources. To address this, employees turn to Large Language Model (LLM)-based chatbots, like ChatGPT, for career advice, though their reliability and personalization remain concerns. This research explores how LLMs, accessible through Azure and OpenAI Services, can enhance employee experiences, focusing on AI-driven professional mentorship.

We investigate benefits, challenges, and strategies to optimize mentorship using advanced technologies. Our methodology comprises three phases: data collection through interviews and platform analysis, LLM utilization for refining AI responses, and developing a Retrieval-Augmented Generation (RAG) chatbot to assess AI mentor effects across industries.

Empirical data collection involves interviews with employees, managers, and mentors, supplemented by platform data, creating a robust dataset. We employ LLMs to answer common queries, iteratively refining responses for accuracy and emotional dynamics.

This project’s potential impact on Employee Experience Management and Human Resource Management is substantial, promising to enrich mentorship experiences and improve HR processes through AI innovation.

Investigating the Utility and Harms of Large Language Models (LLMs) for Mental Health Support 📝 Georgia Institute of Technology : Munmun De Choudhury (PI) In collaboration with mental health clinicians from Northwell Health, we explore the question of where LLM-based chatbots may be useful for mental health contexts, and where they may be harmful.

We conduct audits of how LLM-based chatbots (accessed via the Azure OpenAI Service) respond to pregenerated queries seeking support, and how responses from chatbots compare to how peer supporters might answer the query. We identify where chatbots provide credible mental health information and support, and where they may provide poor advice or propagate misinformation.

This work contributes a beginning framework around the harms of generative AI for mental health, including methods for studying generative AI in mental health and ethical considerations.

“It’s a conversation, not a quiz”: A Risk Taxonomy and Reflection Tool for LLM Adoption in Public Health (opens in new tab) Large Language Model Performance in Nuanced Clinical Concept Extraction and Understanding from Critical Care Documentation: An Evaluation by Clinical Experts Emory University : Craig Jabaley (PI) The research proposal aims to evaluate the proficiency of LLMs in extracting and understanding clinical concepts from routine clinical documentation in adult critical care.

The process involves comparing LLM outputs against human-annotated clinical notes. The study seeks to understand the capabilities, strengths, and limitations of LLMs in the realm of adult critical care. Large Language Model Powered Teaching Assistant for Computer Science Courses Emory University : Pedram Rooshenas (PI) This project aims to develop an AI-based teaching assistant by leveraging large language models (LLMs).

Through precise fine-tuning and strategic prompting, this system will be capable of offering constructive feedback to students and responding to their course-specific queries. Moreover, by incorporating feedback from human educators, we steer the LLMs to produce responses that exemplify the thought process essential for mastering each concept in the course.

We are going to pilot our system for Database Systems, a core course in the Computer Science program. Our proposed system has the potential to enhance the learning experience in public universities, particularly in light of the significant rise in enrollments for computer science and data science programs.

Large Language Models and Digital Empathy* 📝 University of Texas at Austin : Desmond Ong (PI) * AICE Accelerator collaboration The Digital Empathy pilot aims to investigate emotional intelligence in Large Foundation Model (LFM) -driven systems and to develop and study a series of empathic AI agents to understand and augment human performance and wellbeing.

Until now, there has been very little empirical evidence of how empathic LFM systems are or the psychological implications of these systems during human-AI interactions. The project will contribute to a comprehensive survey of the research opportunities and priorities concerning empathy in AI systems and a research platform for the systematic evaluation of empathic agents.

Large Language Models Produce Responses Perceived to be Empathic (opens in new tab) Large Language Models are Capable of Offering Cognitive Reappraisal, if Guided (opens in new tab) Large Language Models: By the People, For the People Carnegie Mellon University : Zachary Lipton (PI) The proposal aims to study the statistical and algorithmic foundation of training and applying LLMs in a human-centered manner.

Leveraging Azure OpenAI Services in Adult Learning Support Tools 📝 University of Toronto : Anastasia Kuzminykh (PI) We investigate the use of randomized A/B experiments to provide a more scientific basis for