Malware authors often employ code obfuscations to make their malware harder to detect. Existing tools for generating obfuscated code often require access to the original source code (e.g., C++ or Java), and adding new obfuscations is a non-trivial, labor-intensive process. In this study, we ask the following question: Can Large Language Models (LLMs) potentially generate a new obfuscated assembly code? If so, this poses a risk to anti-virus engines and potentially increases the flexibility of attackers to create new obfuscation patterns. We answer this in the affirmative by developing the MetamorphASM benchmark comprising MetamorphASM Dataset (MAD) along with three code obfuscation techniques: dead code, register substitution, and control flow change. The MetamorphASM systematically evaluates the ability of LLMs to generate and analyze obfuscated code using MAD, which contains 328,200 obfuscated assembly code samples. We release this dataset and analyze the success rate of various LLMs (e.g., GPT-3.5/4, GPT-4o-mini, Starcoder, CodeGemma, CodeLlama, CodeT5, and LLaMA 3.1) in generating obfuscated assembly code. The evaluation was performed using established information-theoretic metrics and manual human review to ensure correctness and provide the foundation for researchers to study and develop remediations to this risk.

Seyedali Mohammadi, Bhaskara Hanuma Vedula, Hemank Lamba, Edward Raff, Ponnurangam Kumaraguru, Francis Ferraro, Manas Gaur. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Providing contextual and comprehensive medical information tailored to individual patients is critical for enabling effective care in the healthcare domain. However, existing approaches often struggle to deliver personalized responses due to the distributed nature of medical data across multiple sources such as patient records, medical literature, and online resources. To address this challenge, we present MedInsight , a multi-source context augmentation framework that leverages Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) techniques to augment patient-specific information from medical transcripts with trusted knowledge from textbooks and web resources to generate personalized and contextually relevant responses. Our framework consists of three phases: patient context retrieval, medical knowledge retrieval, and response generation. By augmenting patient context with relevant external knowledge, MedInsight generates contextually relevant responses, empowering patients and caregivers with actionable insights. Experiments on the MTSamples dataset validate MedInsight ’s effectiveness in generating contextually appropriate medical responses, using a comprehensive set of metrics including RAGAs, TruLens, ROUGE, and BertScore. Additionally, qualitative evaluations by Subject Matter Experts (SMEs) further confirm the relevance and factual correctness of the generated responses.

Dialogue systems are designed to offer human users social support or functional services through natural language interactions. Traditional conversation research has put significant emphasis on a system’s response-ability, including its capacity to understand dialogue context and generate appropriate responses. However, the key element of proactive behavior—a crucial aspect of intelligent conversations—is often overlooked in these studies. Proactivity empowers conversational agents to lead conversations towards achieving pre-defined targets or fulfilling specific goals on the system side. Proactive dialogue systems are equipped with advanced techniques to handle complex tasks, requiring strategic and motivational interactions, thus representing a significant step towards artificial general intelligence. Motivated by the necessity and challenges of building proactive dialogue systems, we provide a comprehensive review of various prominent problems and advanced designs for implementing proactivity into different types of dialogue systems, including open-domain dialogues, task-oriented dialogues, and information-seeking dialogues. We also discuss real-world challenges that require further research attention to meet application needs in the future, such as proactivity in dialogue systems that are based on large language models, proactivity in hybrid dialogues, evaluation protocols and ethical considerations for proactive dialogue systems. By providing a quick access and overall picture of the proactive dialogue systems domain, we aim to inspire new research directions and stimulate further advancements towards achieving the next level of conversational AI capabilities, paving the way for more dynamic and intelligent interactions within various application domains.

The lack of explainability in using relevant clinical knowledge hinders the adoption of artificial intelligence-powered analysis of unstructured clinical dialogue. A wealth of relevant, untapped Mental Health (MH) data is available in online communities, providing the opportunity to address the explainability problem with substantial potential impact as a screening tool for both online and offline applications. Inspired by how clinicians rely on their expertise when interacting with patients, we leverage relevant clinical knowledge to classify and explain depression-related data, reducing manual review time and engendering trust. We developed a method to enhance attention in contemporary transformer models and generate explanations for classifications that are understandable by mental health practitioners (MHPs) by incorporating external clinical knowledge. We propose a domain-general architecture called ProcesS knowledgeinfused cross ATtention (PSAT) that incorporates clinical practice guidelines (CPG) when computing attention. We transform a CPG resource focused on depression, such as the Patient Health Questionnaire (e.g. PHQ-9) and related questions, into a machine-readable ontology using SNOMED-CT. With this resource, PSAT enhances the ability of models like GPT-3.5 to generate application-relevant explanations. Evaluation of four expert-curated datasets related to depression demonstrates PSAT's applicationrelevant explanations. PSAT surpasses the performance of twelve baseline models and can provide explanations where other baselines fall short.

Monitoring public sentiment via social media is potentially helpful during health crises such as the COVID-19 pandemic. However, traditional frequency-based and data-driven neural network-based approaches can miss newly relevant content due to the evolving nature of language in a dynamic environment. Human-curated symbolic knowledge sources, such as lexicons for standard language and slang terms, can potentially elevate social media signals in evolving language. We introduce a neurosymbolic method that integrates neural networks with symbolic knowledge sources, improving the detection and interpretation of mental health-related tweets relevant to COVID-19. Our method was evaluated using a corpus of large datasets (~12 billion tweets, 2.5 million subreddit data, and 700k news articles) and multiple knowledge graphs. This method dynamically adapts to evolving language, outperforming purely data-driven models with an F1 score exceeding 92%. This approach also showed faster adaptation to new data and lower computational demands than fine-tuning pre-trained large language models (LLMs). This study demonstrates the benefit of neurosymbolic methods in interpreting text in a dynamic environment for tasks such as health surveillance.

Responsible AI is built upon a set of principles that prioritize fairness, transparency, accountability, and inclusivity in AI development and deployment. As AI systems become increasingly sophisticated, including the explosion of generative AI, there is a growing need to address ethical considerations and potential societal impacts of their uses. Knowledge graphs (KGs), as structured representations of information, can enhance generative AI performance by providing context, explaining outputs, and reducing biases, thereby offering a powerful framework to address the challenges of responsible AI. By leveraging semantic relationships and contextual understanding, KGs facilitate transparent decision-making, enabling stakeholders to trace and interpret the reasoning behind AI driven outcomes. Moreover, they provide a means to capture and manage diverse knowledge sources, supporting the development of fair and unbiased AI models. The workshop aims to investigate the role of knowledge graphs in promoting responsible AI principles and creating a cooperative space for researchers, practitioners, and policymakers to exchange insights and enhance their comprehension of KGs' impact on achieving responsible AI solutions. It seeks to facilitate collaboration and idea-sharing to advance the understanding of how KGs can contribute to responsible AI.

Despite recent advancements showcasing the impressive capabilities of Large Language Models (LLMs) in conversational systems, we show that even state-of-the-art LLMs are morally inconsistent in their generations, questioning their reliability (and trustworthiness in general). Prior works in LLM evaluation focus on developing ground-truth data to measure accuracy on specific tasks. However, for moral scenarios that often lack universally agreed-upon answers, consistency in model responses becomes crucial for their reliability. To address this issue, we propose an information-theoretic measure called Semantic Graph Entropy (SaGE), grounded in the concept of “Rules of Thumb” (RoTs) to measure a model’s moral consistency. RoTs are abstract principles learned by a model and can help explain their decision-making strategies effectively. To this extent, we construct the Moral Consistency Corpus (MCC), containing 50K moral questions, responses to them by LLMs, and the RoTs that these models followed. Furthermore, to illustrate the generalizability of SaGE, we use it to investigate LLM consistency on two popular datasets – TruthfulQA and HellaSwag. Our results reveal that task accuracy and consistency are independent problems, and there is a dire need to investigate these issues further.

The Knowledge-infused Learning Workshop is a recurring event in ACM's KDD Conference that gathers the research community on knowledge graphs and knowledge-enabled learning, grounded neurosymbolic AI, explainable and safe AI, and applications in highstakes decision-making problems.This year, the workshop aligned with Biden's vision of Responsible AI Development [1].

Conversational Systems (CSys) represent practical and tangible outcomes of advances in NLP and AI. CSys see continuous improvements through unsupervised training of large language models (LLMs) on a humongous amount of generic training data. However, when these CSys are suggested for use in domains like Mental Health, they fail to match the acceptable standards of clinical care, such as the clinical process in Patient Health Questionnaire (PHQ-9). The talk will present, Knowledge-infused Learning (KiL), a paradigm within NeuroSymbolic AI that focuses on making machine/deep learning models (i) learn over knowledge-enriched data, (ii) learn to follow guidelines in process-oriented tasks for safe and reasonable generation, and (iii) learn to leverage multiple contexts and stratified knowledge to yield user-level explanations. KiL established Knowledge-Intensive Language Understanding, a set of tasks for assessing safety, explainability, and conceptual flow in CSys.

The Natural Language Processing (NLP) community has been using crowd-sourcing techniques to create benchmark datasets such as General Language Understanding and Evaluation (GLUE) for training modern Language Models (LMs) such as BERT. GLUE tasks measure the reliability scores using inter-annotator metrics - Cohen's Kappa (K). However, the reliability aspect of LMs has often been overlooked. To counter this problem, we explore a knowledge-guided LM ensembling approach that leverages reinforcement learning to integrate knowledge from ConceptNet and Wikipedia as knowledge graph embeddings. This approach mimics human annotators resorting to external knowledge to compensate for information deficits in the datasets. Across nine GLUE datasets, our research shows that ensembling strengthens reliability and accuracy scores, outperforming state-of-the-art.

Recent prolific advances in artificial intelligence through the incorporation of domain knowledge have constituted a new paradigm for AI and data mining communities. For example, the human feedback-based language generation in ChatGPT (a large language model (LLM)), the use of Protein Bank in DeepMind's AlphaFold, and the use of 23 rules of safety in DeepMind's Sparrow have demonstrated the success of teaming human knowledge and AI. In addition, the knowledge retrieval-guided language modeling methods have strengthened the association between knowledge and AI. However, translating research methods and resources into practice presents a new challenge for the machine learning and data/knowledge mining communities. For example, in DARPA's Explainable AI seminar, the need for explainable contextual adaptation is seen as the 3rd phase of AI, facilitating the interplay between data and knowledge for explainability, safety, and, eventually, trust. However, policymakers and practitioners assert serious usability and privacy concerns that constrain adoption, notably in high-consequence domains, such as cybersecurity, healthcare, and other social good domains. In addition, limitations in output quality, measurement, and interactive ability, including both the provision of explanations and the acceptance of user preferences, result in low adoption rates in such domains. This workshop aims to accelerate our pace towards creating innovative methods for integrating knowledge into contemporary AI and data science methods and develop metrics for assessing performance in various applications.

Low self-esteem and interpersonal needs (i.e., thwarted belongingness (TB) and perceived burdensomeness (PB)) have a major impact on depression and suicide attempts. Individuals seek social connectedness on social media to boost and alleviate their loneliness. Social media platforms allow people to express their thoughts, experiences, beliefs, and emotions. Prior studies on mental health from social media have focused on symptoms, causes, and disorders. Whereas an initial screening of social media content for interpersonal risk factors and low self-esteem may raise early alerts and assign therapists to at-risk users of mental disturbance. Standardized scales measure self-esteem and interpersonal needs from questions created using psychological theories. In the current research, we introduce a psychology-grounded and expertly annotated dataset, LoST: Low Self esTeem, to study and detect low self-esteem on Reddit. Through an annotation approach involving checks on coherence, correctness, consistency, and reliability, we ensure gold-standard for supervised learning. We present results from different deep language models tested using two data augmentation techniques. Our findings suggest developing a class of language models that infuses psychological and clinical knowledge.

Conversational Information Seeking (CIS) is a relatively new research area within conversational AI that attempts to seek information from end-users in order to understand and satisfy the users' needs. If realized, such a CIS system has far-reaching benefits in the real world; for example, CIS systems can assist clinicians in pre-screening or triaging patients in healthcare. A key open sub-problem in CIS that remains unaddressed in the literature is generating Information Seeking Questions (ISQs) based on a short initial query from the end-user. To address this open problem, we propose Information SEEking Question generator (ISEEQ), a novel approach for generating ISQs from just a short user query, given a large text corpus relevant to the user query. Firstly, ISEEQ uses a knowledge graph to enrich the user query. Secondly, ISEEQ uses the knowledge-enriched query to retrieve relevant context passages to ask coherent ISQs adhering to a conceptual flow. Thirdly, ISEEQ introduces a new deep generative-adversarial reinforcement learning-based approach for generating ISQs. We show that ISEEQ can generate high-quality ISQs to promote the development of CIS agents. ISEEQ significantly outperforms comparable baselines on five ISQ evaluation metrics across four datasets having user queries from diverse domains. Further, we argue that ISEEQ is transferable across domains for generating ISQs, as it shows the acceptable performance when trained and tested on different pairs of domains. A qualitative human evaluation confirms that ISEEQ generated ISQs are comparable in quality to human-generated questions, and it outperformed the best comparable baseline.

Social media plays a major role during and after major natural disasters (e.g., hurricanes, large-scale fires, etc.), as people “on the ground” post useful information on what is actually happening. Given the large amounts of posts, a major challenge is identifying the information that is useful and actionable. Emergency responders are largely interested in finding out what events are taking place so they can properly plan and deploy resources. In this paper we address the problem of automatically identifying important sub-events (within a large-scale emergency “event”, such as a hurricane). In particular, we present a novel, unsupervised learning framework to detect sub-events in Tweets for retrospective crisis analysis. We first extract noun-verb pairs and phrases from raw tweets as sub-event candidates. Then, we learn a semantic embedding of extracted noun-verb pairs and phrases, and rank them against a crisis-specific ontology. We filter out noisy and irrelevant information then cluster the noun-verb pairs and phrases so that the top-ranked ones describe the most important sub-events. Through quantitative experiments on two large crisis data sets (Hurricane Harvey and the 2015 Nepal Earthquake), we demonstrate the effectiveness of our approach over the state-of-the-art. Our qualitative evaluation shows better performance compared to our baseline.

Mental health illness such as depression is a significant risk factor for suicide ideation, behaviors, and attempts. A report by Substance Abuse and Mental Health Services Administration (SAMHSA) shows that 80% of the patients suffering from Borderline Personality Disorder (BPD) have suicidal behavior, 5-10% of whom commit suicide. While multiple initiatives have been developed and implemented for suicide prevention, a key challenge has been the social stigma associated with mental disorders, which deters patients from seeking help or sharing their experiences directly with others including clinicians. This is particularly true for teenagers and younger adults where suicide is the second highest cause of death in the US. Prior research involving surveys and questionnaires (e.g. PHQ-9) for suicide risk prediction failed to provide a quantitative assessment of risk that informed timely clinical decision-making for intervention. Our interdisciplinary study concerns the use of Reddit as an unobtrusive data source for gleaning information about suicidal tendencies and other related mental health conditions afflicting depressed users. We provide details of our learning framework that incorporates domain-specific knowledge to predict the severity of suicide risk for an individual. Our approach involves developing a suicide risk severity lexicon using medical knowledge bases and suicide ontology to detect cues relevant to suicidal thoughts and actions. We also use language modeling, medical entity recognition and normalization and negation detection to create a dataset of 2181 redditors that have discussed or implied suicidal ideation, behavior, or attempt. Given the importance of clinical knowledge, our gold standard dataset of 500 redditors (out of 2181) was developed by four practicing psychiatrists following the guidelines outlined in Columbia Suicide Severity Rating Scale (C-SSRS), with the pairwise annotator agreement of 0.79 and group-wise agreement of 0.73. Compared to the existing four-label classification scheme (no risk, low risk, moderate risk, and high risk), our proposed C-SSRS-based 5-label classification scheme distinguishes people who are supportive, from those who show different severity of suicidal tendency. Our 5-label classification scheme outperforms the state-of-the-art schemes by improving the graded recall by 4.2% and reducing the perceived risk measure by 12.5%. Convolutional neural network (CNN) provided the best performance in our scheme due to the discriminative features and use of domain-specific knowledge resources, in comparison to SVM-L that has been used in the state-of-the-art tools over similar dataset.

Terror attacks have been linked in part to online extremist content. Online conversations are cloaked in religious ambiguity, with deceptive intentions, often twisted from mainstream meaning to serve a malevolent ideology. Although tens of thousands of Islamist extremism supporters consume such content, they are a small fraction relative to peaceful Muslims. The efforts to contain the ever-evolving extremism on social media platforms have remained inadequate and mostly ineffective. Divergent extremist and mainstream contexts challenge machine interpretation, with a particular threat to the precision of classification algorithms. Radicalization is a subtle long-running persuasive process that occurs over time. Our context-aware computational approach to the analysis of extremist content on Twitter breaks down this persuasion process into building blocks that acknowledge inherent ambiguity and sparsity that likely challenge both manual and automated classification. Based on prior empirical and qualitative research in social sciences, particularly political science, we model this process using a combination of three contextual dimensions -- religion, ideology, and hate -- each elucidating a degree of radicalization and highlighting independent features to render them computationally accessible. We utilize domain-specific knowledge resources for each of these contextual dimensions such as Qur'an for religion, the books of extremist ideologues and preachers for political ideology and a social media hate speech corpus for hate. The significant sensitivity of the Islamist extremist ideology and its local and global security implications require reliable algorithms for modelling such communications on Twitter. Our study makes three contributions to reliable analysis: (i) Development of a computational approach rooted in the contextual dimensions of religion, ideology, and hate, which reflects strategies employed by online Islamist extremist groups, (ii) An in-depth analysis of relevant tweet datasets with respect to these dimensions to exclude likely mislabeled users, and (iii) A framework for understanding online radicalization as a process to assist counter-programming. Given the potentially significant social impact, we evaluate the performance of our algorithms to minimize mislabeling, where our context-aware approach outperforms a competitive baseline by 10.2% in precision, thereby enhancing the potential of such tools for use in human review.

Social media platforms are increasingly being used to share and seek advice on mental health issues. In particular, Reddit users freely discuss such issues on various subreddits, whose structure and content can be leveraged to formally interpret and relate subreddits and their posts in terms of mental health diagnostic categories. There is prior research on the extraction of mental health-related information, including symptoms, diagnosis, and treatments from social media; however, our approach can additionally provide actionable information to clinicians about the mental health of a patient in diagnostic terms for web-based intervention. Specifically, we provide a detailed analysis of the nature of subreddit content from domain expert's perspective and introduce a novel approach to map each subreddit to the best matching DSM-5 (Diagnostic and Statistical Manual of Mental Disorders - 5th Edition) category using multi-class classifier. Our classification algorithm analyzes all the posts of a subreddit by adapting topic modeling and word-embedding techniques, and utilizing curated medical knowledge bases to quantify relationship to DSM-5 categories. Our semantic encoding-decoding optimization approach reduces the false-alarm-rate from 30% to 2.5% over a comparable heuristic baseline, and our mapping results have been verified by domain experts achieving a kappa score of 0.84.

We partner with a leading European healthcare provider and design a mechanism to match patients with family doctors in primary care. We define the matchmaking process for several distinct use cases given different levels of available information about patients. Then, we adopt a hybrid recommender system to present each patient a list of family doctor recommendations. In particular, we model patient trust of family doctors using a large-scale dataset of consultation histories, while accounting for the temporal dynamics of their relationships. Our proposed approach shows higher predictive accuracy than both a heuristic baseline and a collaborative filtering approach, and the proposed trust measure further improves model performance.

A crowd sampled from a set of individuals can provide a more accurate prediction in aggregate than most individuals.This effect, referred to as wisdom of crowd, exists when crowd members bring diverse perspectives to decision making. Such diversity leads to uncorrelated prediction errors that cancel out in aggregate. As crowd members' judgments are often the result of solution strategies, diversity in solution strategies can enhance crowd wisdom. One of the most challenging tasks in sampling such a crowd is to determine the individual's solution strategy for a prediction problem. As participating individuals often share their perspectives through social media, we can use such data to identify an individual's solution strategy. In this paper, we propose a crowd selection approach using social media posts (tweets) indicating diverse solution strategies. We use tweet classification to identify participants' prediction strategies and categorize participants based on the binomial test to identify sets of participants that apply a similar strategy. We then form a diverse crowd by sampling participants from different sets. Using the domain of Fantasy Sports, we show that such a diverse crowd can outperform crowd selected at random and 90% of individual participants, and participant categorization schemes using word2vec. Further, we use a knowledge graph to investigate the factors forming such a diverse crowd and how these factors can lead to a better decision. Relative to bottom-up (data-driven) processes the approach presented here provides an explanation of diverse crowd behavior.

With 93% of pro-marijuana population in US favoring legalization of medical marijuana, high expectations of a greater return for Marijuana stocks, and public actively sharing information about medical, recreational and business aspects related to marijuana, it is no surprise that marijuana culture is thriving on Twitter. After the legalization of marijuana for recreational and medical purposes in 29 states, there has been a dramatic increase in the volume of drug-related communications on Twitter. Specifically, Twitter accounts have been established for promotional and informational purposes, some prominent among them being American Ganja, Medical Marijuana Exchange, and Cannabis Now. Identification and characterization of different user types can allow us to conduct more fine-grained spatiotemporal analysis to identify dominant or emerging topics in the echo chambers of marijuana-related communities on Twitter. In this research, we mainly focus on classifying Twitter accounts created and run by ordinary users, retailers, and informed agencies. Classifying user accounts by type can enable better capturing and highlighting of aspects such as trending topics, business profiling of marijuana companies, and state-specific marijuana policymaking. Furthermore, type-based analysis can provide more profound understanding and reliable assessment of the implications of marijuana-related communications. We developed a comprehensive approach to classifying users by their types on Twitter through contextualization of their marijuana-related conversations. We accomplished this using compositional multiview embedding synthesized from People, Content, and Network views achieving 8% improvement over the empirical baseline.