Mayank Kejriwal

Henrique Santos, Ke Shen, Alice M. Mulvehill et al. · stanford

Programming machines with commonsense reasoning (CSR) abilities is a longstanding challenge in the Artificial Intelligence community. Current CSR benchmarks use multiple-choice (and in relatively fewer cases, generative) question-answering instances to evaluate machine commonsense. Recent progress in transformer-based language representation models suggest that considerable progress has been made on existing benchmarks. However, although tens of CSR benchmarks currently exist, and are growing, it is not evident that the full suite of commonsense capabilities have been systematically evaluated. Furthermore, there are doubts about whether language models are 'fitting' to a benchmark dataset's training partition by picking up on subtle, but normatively irrelevant (at least for CSR), statistical features to achieve good performance on the testing partition. To address these challenges, we propose a benchmark called Theoretically-Grounded Commonsense Reasoning (TG-CSR) that is also based on discriminative question answering, but with questions designed to evaluate diverse aspects of commonsense, such as space, time, and world states. TG-CSR is based on a subset of commonsense categories first proposed as a viable theory of commonsense by Gordon and Hobbs. The benchmark is also designed to be few-shot (and in the future, zero-shot), with only a few training and validation examples provided. This report discusses the structure and construction of the benchmark. Preliminary results suggest that the benchmark is challenging even for advanced language representation models designed for discriminative CSR question answering tasks. Benchmark access and leaderboard: https://codalab.lisn.upsaclay.fr/competitions/3080 Benchmark website: https://usc-isi-i2.github.io/TGCSR/

Ashwath Vaithinathan Aravindan, Mayank Kejriwal

Chain-of-Thought (CoT) prompting has emerged as a foundational technique for eliciting reasoning from Large Language Models (LLMs), yet the robustness of this approach to corruptions in intermediate reasoning steps remains poorly understood. This paper presents a comprehensive empirical evaluation of LLM robustness to a structured taxonomy of 5 CoT perturbation types: \textit{MathError, UnitConversion, Sycophancy, SkippedSteps,} and \textit{ExtraSteps}. We evaluate 13 models spanning three orders of magnitude in parameter count, testing their ability to complete mathematical reasoning tasks despite perturbations injected in the reasoning chain. Our key findings reveal heterogeneous vulnerability patterns: MathError perturbations produce the most severe degradation in small models (50-60\% accuracy loss) but show strong scaling benefits; UnitConversion remains challenging across all scales (>5\% loss even for midsized models); ExtraSteps incur minimal accuracy degradation (0-6\%) even for the smallest of models; Sycophancy and SkippedSteps produce modest effects ($\sim$10\% loss for small models) and slightly improve with scale. Scaling relationships show that model size serve as a protective factor against many perturbations but not always. These findings have direct implications for deploying LLMs in multi-stage reasoning pipelines and underscore the necessity of task-specific robustness assessments and mitigation strategies. The code and results are available at https://github.com/Mystic-Slice/CoTPerturbation

Ashwath Vaithinathan Aravindan, Mayank Kejriwal

Large language models often generate confident but incorrect answers rather than abstaining when uncertain. This problem is particularly acute for small language models (SLMs), where computational constraints and autonomous operation amplify the need for reliable uncertainty detection. We propose _Second Guess_, a lightweight, parameter-free prompting technique for abstention in multiple-choice question answering (MCQA) that is well-suited for SLMs. Our key empirical insight is that models which truly know an answer will select it consistently, while uncertain models exhibit unstable behavior when an ``I don't know'' option is added. Evaluated on four open models (2B-8B parameters) and four benchmarks, Second Guess achieves the highest composite risk improvement of 10.81\%. Notably, it maintains an 8\% composite risk improvement on fine-tuned models where entropy-based methods degrade, and improves most for lower-performing models. All code and results required to reproduce this work is available in https://github.com/Mystic-Slice/second-guess

Ke Shen, Mayank Kejriwal

Despite their impressive performance, large language models (LLMs) such as ChatGPT are known to pose important risks. One such set of risks arises from misplaced confidence, whether over-confidence or under-confidence, that the models have in their inference. While the former is well studied, the latter is not, leading to an asymmetry in understanding the comprehensive risk of the model based on misplaced confidence. In this paper, we address this asymmetry by defining two types of risk (decision and composite risk), and proposing an experimental framework consisting of a two-level inference architecture and appropriate metrics for measuring such risks in both discriminative and generative LLMs. The first level relies on a decision rule that determines whether the underlying language model should abstain from inference. The second level (which applies if the model does not abstain) is the model's inference. Detailed experiments on four natural language commonsense reasoning datasets using both an open-source ensemble-based RoBERTa model and ChatGPT, demonstrate the practical utility of the evaluation framework. For example, our results show that our framework can get an LLM to confidently respond to an extra 20.1% of low-risk inference tasks that other methods might misclassify as high-risk, and skip 19.8% of high-risk tasks, which would have been answered incorrectly.

Navapat Nananukul, Khanin Sisaengsuwanchai, Mayank Kejriwal

Entity Resolution (ER) is the problem of semi-automatically determining when two entities refer to the same underlying entity, with applications ranging from healthcare to e-commerce. Traditional ER solutions required considerable manual expertise, including domain-specific feature engineering, as well as identification and curation of training data. Recently released large language models (LLMs) provide an opportunity to make ER more seamless and domain-independent. However, it is also well known that LLMs can pose risks, and that the quality of their outputs can depend on how prompts are engineered. Unfortunately, a systematic experimental study on the effects of different prompting methods for addressing unsupervised ER, using LLMs like ChatGPT, has been lacking thus far. This paper aims to address this gap by conducting such a study. We consider some relatively simple and cost-efficient ER prompt engineering methods and apply them to ER on two real-world datasets widely used in the community. We use an extensive set of experimental results to show that an LLM like GPT3.5 is viable for high-performing unsupervised ER, and interestingly, that more complicated and detailed (and hence, expensive) prompting methods do not necessarily outperform simpler approaches. We provide brief discussions on qualitative and error analysis, including a study of the inter-consistency of different prompting methods to determine whether they yield stable outputs. Finally, we consider some limitations of LLMs when applied to ER.

Katarina Doctor, Christine Task, Eric Kildebeck et al.

Artificial Intelligence (AI) systems planned for deployment in real-world applications frequently are researched and developed in closed simulation environments where all variables are controlled and known to the simulator or labeled benchmark datasets are used. Transition from these simulators, testbeds, and benchmark datasets to more open-world domains poses significant challenges to AI systems, including significant increases in the complexity of the domain and the inclusion of real-world novelties; the open-world environment contains numerous out-of-distribution elements that are not part in the AI systems' training set. Here, we propose a path to a general, domain-independent measure of domain complexity level. We distinguish two aspects of domain complexity: intrinsic and extrinsic. The intrinsic domain complexity is the complexity that exists by itself without any action or interaction from an AI agent performing a task on that domain. This is an agent-independent aspect of the domain complexity. The extrinsic domain complexity is agent- and task-dependent. Intrinsic and extrinsic elements combined capture the overall complexity of the domain. We frame the components that define and impact domain complexity levels in a domain-independent light. Domain-independent measures of complexity could enable quantitative predictions of the difficulty posed to AI systems when transitioning from one testbed or environment to another, when facing out-of-distribution data in open-world tasks, and when navigating the rapidly expanding solution and search spaces encountered in open-world domains.

Zhisheng Tang, Mayank Kejriwal

Spatial reasoning, an important faculty of human cognition with many practical applications, is one of the core commonsense skills that is not purely language-based and, for satisfying (as opposed to optimal) solutions, requires some minimum degree of planning. Existing benchmarks of Commonsense Spatial Reasoning (CSR) tend to evaluate how Large Language Models (LLMs) interpret text-based spatial $\textit{descriptions}$ rather than directly evaluate a plan produced by the LLM in response to a $\textit{specific}$ spatial reasoning problem. In this paper, we construct a large-scale benchmark called GRASP, which consists of 16,000 grid-based environments where the agent is tasked with an energy collection problem. These environments include 100 grid instances instantiated using each of the 160 different grid settings, involving five different energy distributions, two modes of agent starting position, and two distinct obstacle configurations, as well as three kinds of agent constraints. Using GRASP, we compare classic baseline approaches, such as random walk and greedy search methods, with advanced LLMs like GPT-3.5-Turbo, GPT-4o, and GPT-o1-mini. The experimental results indicate that even these advanced LLMs struggle to consistently achieve satisfactory solutions.

Zhisheng Tang, Mayank Kejriwal

In recent years, transformer-based language representation models (LRMs) have achieved state-of-the-art results on difficult natural language understanding problems, such as question answering and text summarization. As these models are integrated into real-world applications, evaluating their ability to make rational decisions is an important research agenda, with practical ramifications. This article investigates LRMs' rational decision-making ability through a carefully designed set of decision-making benchmarks and experiments. Inspired by classic work in cognitive science, we model the decision-making problem as a bet. We then investigate an LRM's ability to choose outcomes that have optimal, or at minimum, positive expected gain. Through a robust body of experiments on four established LRMs, we show that a model is only able to `think in bets' if it is first fine-tuned on bet questions with an identical structure. Modifying the bet question's structure, while still retaining its fundamental characteristics, decreases an LRM's performance by more than 25\%, on average, although absolute performance remains well above random. LRMs are also found to be more rational when selecting outcomes with non-negative expected gain, rather than optimal or strictly positive expected gain. Our results suggest that LRMs could potentially be applied to tasks that rely on cognitive decision-making skills, but that more research is necessary before they can robustly make rational decisions.

Ke Shen, Mayank Kejriwal

In recent years,Text-to-SQL, the problem of automatically converting questions posed in natural language to formal SQL queries, has emerged as an important problem at the intersection of natural language processing and data management research. Large language models (LLMs) have delivered impressive performance when used in an off-the-shelf performance, but still fall significantly short of expected expert-level performance. Errors are especially probable when a nuanced understanding is needed of database schemas, questions, and SQL clauses to do proper Text-to-SQL conversion. We introduce SelECT-SQL, a novel in-context learning solution that uses an algorithmic combination of chain-of-thought (CoT) prompting, self-correction, and ensemble methods to yield a new state-of-the-art result on challenging Text-to-SQL benchmarks. Specifically, when configured using GPT-3.5-Turbo as the base LLM, SelECT-SQL achieves 84.2% execution accuracy on the Spider leaderboard's development set, exceeding both the best results of other baseline GPT-3.5-Turbo-based solutions (81.1%), and the peak performance (83.5%) of the GPT-4 result reported on the leaderboard.

Zhisheng Tang, Mayank Kejriwal

We conduct a pilot study selectively evaluating the cognitive abilities (decision making and spatial reasoning) of two recently released generative transformer models, ChatGPT and DALL-E 2. Input prompts were constructed following neutral a priori guidelines, rather than adversarial intent. Post hoc qualitative analysis of the outputs shows that DALL-E 2 is able to generate at least one correct image for each spatial reasoning prompt, but most images generated are incorrect (even though the model seems to have a clear understanding of the objects mentioned in the prompt). Similarly, in evaluating ChatGPT on the rationality axioms developed under the classical Von Neumann-Morgenstern utility theorem, we find that, although it demonstrates some level of rational decision-making, many of its decisions violate at least one of the axioms even under reasonable constructions of preferences, bets, and decision-making prompts. ChatGPT's outputs on such problems generally tended to be unpredictable: even as it made irrational decisions (or employed an incorrect reasoning process) for some simpler decision-making problems, it was able to draw correct conclusions for more complex bet structures. We briefly comment on the nuances and challenges involved in scaling up such a 'cognitive' evaluation or conducting it with a closed set of answer keys ('ground truth'), given that these models are inherently generative and open-ended in responding to prompts.

Ke Shen, Mayank Kejriwal

Recent work on transformer-based neural networks has led to impressive advances on multiple-choice natural language understanding (NLU) problems, such as Question Answering (QA) and abductive reasoning. Despite these advances, there is limited work still on understanding whether these models respond to perturbed multiple-choice instances in a sufficiently robust manner that would allow them to be trusted in real-world situations. We present four confusion probes, inspired by similar phenomena first identified in the behavioral science community, to test for problems such as prior bias and choice paralysis. Experimentally, we probe a widely used transformer-based multiple-choice NLU system using four established benchmark datasets. Here we show that the model exhibits significant prior bias and to a lesser, but still highly significant degree, choice paralysis, in addition to other problems. Our results suggest that stronger testing protocols and additional benchmarks may be necessary before the language models are used in front-facing systems or decision making with real world consequences.

Ke Shen, Mayank Kejriwal

Large Language Models (LLMs), such as ChatGPT, have achieved impressive milestones in natural language processing (NLP). Despite their impressive performance, the models are known to pose important risks. As these models are deployed in real-world applications, a systematic understanding of different risks posed by these models on tasks such as natural language inference (NLI), is much needed. In this paper, we define and formalize two distinct types of risk: decision risk and composite risk. We also propose a risk-centric evaluation framework, and four novel metrics, for assessing LLMs on these risks in both in-domain and out-of-domain settings. Finally, we propose a risk-adjusted calibration method called DwD for helping LLMs minimize these risks in an overall NLI architecture. Detailed experiments, using four NLI benchmarks, three baselines and two LLMs, including ChatGPT, show both the practical utility of the evaluation framework, and the efficacy of DwD in reducing decision and composite risk. For instance, when using DwD, an underlying LLM is able to address an extra 20.1% of low-risk inference tasks (but which the LLM erroneously deems high-risk without risk adjustment) and skip a further 19.8% of high-risk tasks, which would have been answered incorrectly.

Mayank Kejriwal

Entity Resolution (ER) is the problem of determining when two entities refer to the same underlying entity. The problem has been studied for over 50 years, and most recently, has taken on new importance in an era of large, heterogeneous 'knowledge graphs' published on the Web and used widely in domains as wide ranging as social media, e-commerce and search. This chapter will discuss the specific problem of named ER in the context of personal knowledge graphs (PKGs). We begin with a formal definition of the problem, and the components necessary for doing high-quality and efficient ER. We also discuss some challenges that are expected to arise for Web-scale data. Next, we provide a brief literature review, with a special focus on how existing techniques can potentially apply to PKGs. We conclude the chapter by covering some applications, as well as promising directions for future research.

Ke Shen, Mayank Kejriwal

Acquiring commonsense knowledge and reasoning is an important goal in modern NLP research. Despite much progress, there is still a lack of understanding (especially at scale) of the nature of commonsense knowledge itself. A potential source of structured commonsense knowledge that could be used to derive insights is ConceptNet. In particular, ConceptNet contains several coarse-grained relations, including HasContext, FormOf and SymbolOf, which can prove invaluable in understanding broad, but critically important, commonsense notions such as 'context'. In this article, we present a methodology based on unsupervised knowledge graph representation learning and clustering to reveal and study substructures in three heavily used commonsense relations in ConceptNet. Our results show that, despite having an 'official' definition in ConceptNet, many of these commonsense relations exhibit considerable sub-structure. In the future, therefore, such relations could be sub-divided into other relations with more refined definitions. We also supplement our core study with visualizations and qualitative analyses.

Navapat Nananukul, Mayank Kejriwal

Clinical diagnosis requires answers that are accurate, verifiable, and explicitly grounded in official guidelines. While large language models excel at natural language processing, their tendency to hallucinate undermines their utility in high-stakes medical contexts where precision is essential. Existing retrieval-augmented generation (RAG) systems treat all evidence equally, producing noisy context and generic answers misaligned with clinical practice. We present ClinicBot, an AI system that translates guideline recommendations into trustworthy clinical support through three key advances: (1) structured extraction of clinical guidelines into semantic units (recommendations, tables, definitions, narrative) with explicit provenance, (2) evidence prioritization that ranks content by clinical significance and guideline structure rather than textual similarity, and (3) a web-based interface that presents concise, actionable answers with verifiable evidence. We will demonstrate ClinicBot using diabetes questions from real patients and an additional diabetes risk assessment tool that is faithful to the American Diabetes Association (ADA) Standards of Care in Diabetes (2025). The demonstration will illustrate how semantic knowledge extraction and hierarchical evidence ranking can reliably operate in a multi-agent setting to process complex clinical guidelines at scale.

Vishal Patil, Shree Vaishnavi Bacha, Revanth Yamani et al.

Customer-provided reviews have become an important source of information for business owners and other customers alike. However, effectively analyzing millions of unstructured reviews remains challenging. While large language models (LLMs) show promise for natural language understanding, their application to large-scale review analysis has been limited by computational costs and scalability concerns. This study proposes a hybrid approach that uses LLMs for aspect identification while employing classic machine-learning methods for sentiment classification at scale. Using ChatGPT to analyze sampled restaurant reviews, we identified key aspects of dining experiences and developed sentiment classifiers using human-labeled reviews, which we subsequently applied to 4.7 million reviews collected over 17 years from a major online platform. Regression analysis reveals that our machine-labeled aspects significantly explain variance in overall restaurant ratings across different aspects of dining experiences, cuisines, and geographical regions. Our findings demonstrate that combining LLMs with traditional machine learning approaches can effectively automate aspect-based sentiment analysis of large-scale customer feedback, suggesting a practical framework for both researchers and practitioners in the hospitality industry and potentially, other service sectors.

Mayank Kejriwal, Hamid Haidarian, Min-Hsueh Chiu et al.

Efficiently finding doctors and locations is an important search problem for patients in the healthcare domain, for which traditional information retrieval methods tend not to work optimally. In the last ten years, knowledge graphs (KGs) have emerged as a powerful way to combine the benefits of gleaning insights from semi-structured data using semantic modeling, natural language processing techniques like information extraction, and robust querying using structured query languages like SPARQL and Cypher. In this short paper, we present a KG-based search engine architecture for robustly finding doctors and locations in the healthcare domain. Early results demonstrate that our approach can lead to significantly higher coverage for complex queries without degrading quality.

Zhisheng Tang, Mayank Kejriwal

Research on emergent patterns in Large Language Models (LLMs) has gained significant traction in both psychology and artificial intelligence, motivating the need for a comprehensive review that offers a synthesis of this complex landscape. In this article, we systematically review LLMs' capabilities across three important cognitive domains: decision-making biases, reasoning, and creativity. We use empirical studies drawing on established psychological tests and compare LLMs' performance to human benchmarks. On decision-making, our synthesis reveals that while LLMs demonstrate several human-like biases, some biases observed in humans are absent, indicating cognitive patterns that only partially align with human decision-making. On reasoning, advanced LLMs like GPT-4 exhibit deliberative reasoning akin to human System-2 thinking, while smaller models fall short of human-level performance. A distinct dichotomy emerges in creativity: while LLMs excel in language-based creative tasks, such as storytelling, they struggle with divergent thinking tasks that require real-world context. Nonetheless, studies suggest that LLMs hold considerable potential as collaborators, augmenting creativity in human-machine problem-solving settings. Discussing key limitations, we also offer guidance for future research in areas such as memory, attention, and open-source model development.

Sara Melotte, Mayank Kejriwal

Although the recent rise and uptake of COVID-19 vaccines in the United States has been encouraging, there continues to be significant vaccine hesitancy in various geographic and demographic clusters of the adult population. Surveys, such as the one conducted by Gallup over the past year, can be useful in determining vaccine hesitancy, but can be expensive to conduct and do not provide real-time data. At the same time, the advent of social media suggests that it may be possible to get vaccine hesitancy signals at an aggregate level (such as at the level of zip codes) by using machine learning models and socioeconomic (and other) features from publicly available sources. It is an open question at present whether such an endeavor is feasible, and how it compares to baselines that only use constant priors. To our knowledge, a proper methodology and evaluation results using real data has also not been presented. In this article, we present such a methodology and experimental study, using publicly available Twitter data collected over the last year. Our goal is not to devise novel machine learning algorithms, but to evaluate existing and established models in a comparative framework. We show that the best models significantly outperform constant priors, and can be set up using open-source tools.

Mayank Kejriwal, Jiayuan Ding, Runqi Shao et al.

In this paper, we describe and study the indicator mining problem in the online sex advertising domain. We present an in-development system, FlagIt (Flexible and adaptive generation of Indicators from text), which combines the benefits of both a lightweight expert system and classical semi-supervision (heuristic re-labeling) with recently released state-of-the-art unsupervised text embeddings to tag millions of sentences with indicators that are highly correlated with human trafficking. The FlagIt technology stack is open source. On preliminary evaluations involving five indicators, FlagIt illustrates promising performance compared to several alternatives. The system is being actively developed, refined and integrated into a domain-specific search system used by over 200 law enforcement agencies to combat human trafficking, and is being aggressively extended to mine at least six more indicators with minimal programming effort. FlagIt is a good example of a system that operates in limited label settings, and that requires creative combinations of established machine learning techniques to produce outputs that could be used by real-world non-technical analysts.

Mayank Kejriwal

Word embeddings have made enormous inroads in recent years in a wide variety of text mining applications. In this paper, we explore a word embedding-based architecture for predicting the relevance of a role between two financial entities within the context of natural language sentences. In this extended abstract, we propose a pooled approach that uses a collection of sentences to train word embeddings using the skip-gram word2vec architecture. We use the word embeddings to obtain context vectors that are assigned one or more labels based on manual annotations. We train a machine learning classifier using the labeled context vectors, and use the trained classifier to predict contextual role relevance on test data. Our approach serves as a good minimal-expertise baseline for the task as it is simple and intuitive, uses open-source modules, requires little feature crafting effort and performs well across roles.

Zhisheng Tang, Mayank Kejriwal

Research funding discovery remains fundamentally fragmented: researchers navigate disparate agency portals (e.g., in the United States, NSF, NIH, DARPA, Grants.gov, and many others) with heterogeneous interfaces, search capabilities, and data schemas. We present a compound AI system that unifies this landscape through two tightly coupled components: (1) an aggregation layer that autonomously collects, normalizes, and indexes almost 12,000 federal and nonprofit opportunities from fragmented sources via LLM-equipped browser agents, maintaining a biweekly-updated unified database; and (2) an agentic ReAct-based query processing layer that interprets research context (including from PDF documents) and employs hybrid search combining a structured index with selective web search to retrieve relevant opportunities - while avoiding LLM hallucination. The conversational interface supports iterative refinement through multi-turn interactions, allowing researchers to progressively apply constraints without reformulating their core research description. Results stream in real time with full transparency of intermediate reasoning, enabling appropriate calibration of user trust. Currently used by almost 3,000+ users, our approach demonstrates the feasibility of compound AI in reducing grant discovery time from 30--45 minutes (manual, fragmented portal searches) to under 10 minutes (unified, conversational search).

Cole Gawin, Yidan Sun, Mayank Kejriwal

Large language models (LLMs) have achieved remarkable performance in generating human-like text and solving reasoning tasks of moderate complexity, such as question-answering and mathematical problem-solving. However, their capabilities in tasks requiring deeper cognitive skills, such as common-sense understanding and abstract reasoning, remain under-explored. In this paper, we systematically evaluate abstract common-sense reasoning in LLMs using the ConceptNet knowledge graph. We propose two prompting approaches: instruct prompting, where models predict plausible semantic relationships based on provided definitions, and few-shot prompting, where models identify relations using examples as guidance. Our experiments with the gpt-4o-mini model show that in instruct prompting, consistent performance is obtained when ranking multiple relations but with substantial decline when the model is restricted to predicting only one relation. In few-shot prompting, the model's accuracy improves significantly when selecting from five relations rather than the full set, although with notable bias toward certain relations. These results suggest significant gaps still, even in commercially used LLMs' abstract common-sense reasoning abilities, compared to human-level understanding. However, the findings also highlight the promise of careful prompt engineering, based on selective retrieval, for obtaining better performance.

Navapat Nananukul, Mayank Kejriwal

Recent progress in generative AI, including large language models (LLMs) like ChatGPT, has opened up significant opportunities in fields ranging from natural language processing to knowledge discovery and data mining. However, there is also a growing awareness that the models can be prone to problems such as making information up or `hallucinations', and faulty reasoning on seemingly simple problems. Because of the popularity of models like ChatGPT, both academic scholars and citizen scientists have documented hallucinations of several different types and severity. Despite this body of work, a formal model for describing and representing these hallucinations (with relevant meta-data) at a fine-grained level, is still lacking. In this paper, we address this gap by presenting the Hallucination Ontology or HALO, a formal, extensible ontology written in OWL that currently offers support for six different types of hallucinations known to arise in LLMs, along with support for provenance and experimental metadata. We also collect and publish a dataset containing hallucinations that we inductively gathered across multiple independent Web sources, and show that HALO can be successfully used to model this dataset and answer competency questions.

Ashwath Vaithinathan Aravindan, Zhisheng Tang, Mayank Kejriwal

We propose an iterative programmatic planning (IPP) framework for solving grid-based tasks by synthesizing interpretable agent policies expressed in code using large language models (LLMs). Instead of relying on traditional search or reinforcement learning, our approach uses code generation as policy synthesis, where the LLM outputs executable programs that map environment states to action sequences. Our proposed architecture incorporates several prompting strategies, including direct code generation, pseudocode-conditioned refinement, and curriculum-based prompting, but also includes an iterative refinement mechanism that updates code based on task performance feedback. We evaluate our approach using six leading LLMs and two challenging grid-based benchmarks (GRASP and MiniGrid). Our IPP framework demonstrates improvements over direct code generation ranging from 10\% to as much as 10x across five of the six models and establishes a new state-of-the-art result for GRASP. IPP is found to significantly outperform direct elicitation of a solution from GPT-o3-mini (by 63\% on MiniGrid to 116\% on GRASP), demonstrating the viability of the overall approach. Computational costs of all code generation approaches are similar. While code generation has a higher initial prompting cost compared to direct solution elicitation (\$0.08 per task vs. \$0.002 per instance for GPT-o3-mini), the code can be reused for any number of instances, making the amortized cost significantly lower (by 400x on GPT-o3-mini across the complete GRASP benchmark).

Navapat Nananukul, Mayank Kejriwal

Understanding the landscape of artificial intelligence (AI) and machine learning (ML) adoption across the National Institutes of Health (NIH) portfolio is critical for research funding strategy, institutional planning, and health policy. The advent of large language models (LLMs) has fundamentally transformed research landscape analysis, enabling researchers to perform large-scale semantic extraction from thousands of unstructured research documents. In this paper, we illustrate a human-in-the-loop research methodology for LLMs to automatically classify and summarize research descriptions at scale. Using our methodology, we present a comprehensive analysis of 58,746 NIH-funded biomedical research projects from 2025. We show that: (1) AI constitutes 15.9% of the NIH portfolio with a 13.4% funding premium, concentrated in discovery, prediction, and data integration across disease domains; (2) a critical research-to-deployment gap exists, with 79% of AI projects remaining in research/development stages while only 14.7% engage in clinical deployment or implementation; and (3) health disparities research is severely underrepresented at just 5.7% of AI-funded work despite its importance to NIH's equity mission. These findings establish a framework for evidence-based policy interventions to align the NIH AI portfolio with health equity goals and strategic research priorities.

Shama Magnur, Mayank Kejriwal

The emergence of large language models (LLMs) represents a significant technological shift within the scientific ecosystem, particularly within the field of artificial intelligence (AI). This paper examines structural changes in the AI research landscape using a dataset of arXiv preprints (cs.AI) from 2021 through 2025. Given the rapid pace of AI development, the preprint ecosystem has become a critical barometer for real-time scientific shifts, often preceding formal peer-reviewed publication by months or years. By employing a multi-stage data collection and enrichment pipeline in conjunction with LLM-based institution classification, we analyze the evolution of publication volumes, author team sizes, and academic--industry collaboration patterns. Our results reveal an unprecedented surge in publication output following the introduction of ChatGPT, with academic institutions continuing to provide the largest volume of research. However, we observe that academic--industry collaboration is still suppressed, as measured by a Normalized Collaboration Index (NCI) that remains significantly below the random-mixing baseline across all major subfields. These findings highlight a continuing institutional divide and suggest that the capital-intensive nature of generative AI research may be reshaping the boundaries of scientific collaboration.

Navapat Nananukul, Yue Zhang, Ryan Lee et al.

High-assurance reasoning, particularly in critical domains such as law and medicine, requires conclusions that are accurate, verifiable, and explicitly grounded in evidence. This reasoning relies on premises codified from rules, statutes, and contracts, inherently involving defeasible or non-monotonic logic due to numerous exceptions, where the introduction of a single fact can invalidate general rules, posing significant challenges. While large language models (LLMs) excel at processing natural language, their capabilities in standard inference tasks do not translate to the rigorous reasoning required over high-assurance text guidelines. Core reasoning challenges within such texts often manifest specific logical structures involving negation, implication, and, most critically, defeasible rules and exceptions. In this paper, we propose a novel neurosymbolically-grounded architecture called LOGicalThought (LogT) that uses an advanced logical language and reasoner in conjunction with an LLM to construct a dual symbolic graph context and logic-based context. These two context representations transform the problem from inference over long-form guidelines into a compact grounded evaluation. Evaluated on four multi-domain benchmarks against four baselines, LogT improves overall performance by 11.84% across all LLMs. Performance improves significantly across all three modes of reasoning: by up to +10.2% on negation, +13.2% on implication, and +5.5% on defeasible reasoning compared to the strongest baseline.

Mayank Kejriwal, Shilpa Thomas

We describe GNOME (Generating Novelty in Open-world Multi-agent Environments), an experimental platform that is designed to test the effectiveness of multi-agent AI systems when faced with \emph{novelty}. GNOME separates the development of AI gameplaying agents with the simulator, allowing \emph{unanticipated} novelty (in essence, novelty that is not subject to model-selection bias). Using a Web GUI, GNOME was recently demonstrated at NeurIPS 2020 using the game of Monopoly to foster an open discussion on AI robustness and the nature of novelty in real-world environments. In this article, we further detail the key elements of the demonstration, and also provide an overview of the experimental design that is being currently used in the DARPA Science of Artificial Intelligence and Learning for Open-World Novelty (SAIL-ON) program to evaluate external teams developing novelty-adaptive gameplaying agents.

Yongyi Ji, Zhisheng Tang, Mayank Kejriwal

Personality, a fundamental aspect of human cognition, contains a range of traits that influence behaviors, thoughts, and emotions. This paper explores the capabilities of large language models (LLMs) in reconstructing these complex cognitive attributes based only on simple descriptions containing socio-demographic and personality type information. Utilizing the HEXACO personality framework, our study examines the consistency of LLMs in recovering and predicting underlying (latent) personality dimensions from simple descriptions. Our experiments reveal a significant degree of consistency in personality reconstruction, although some inconsistencies and biases, such as a tendency to default to positive traits in the absence of explicit information, are also observed. Additionally, socio-demographic factors like age and number of children were found to influence the reconstructed personality dimensions. These findings have implications for building sophisticated agent-based simulacra using LLMs and highlight the need for further research on robust personality generation in LLMs.

Katarina Doctor, Mayank Kejriwal, Lawrence Holder et al.

Artificial Intelligence (AI) systems, trained in controlled environments, often struggle in real-world complexities. We propose a general framework for estimating domain complexity across diverse environments, like open-world learning and real-world applications. This framework distinguishes between intrinsic complexity (inherent to the domain) and extrinsic complexity (dependent on the AI agent). By analyzing dimensionality, sparsity, and diversity within these categories, we offer a comprehensive view of domain challenges. This approach enables quantitative predictions of AI difficulty during environment transitions, avoids bias in novel situations, and helps navigate the vast search spaces of open-world domains.

Yuesheng Luo, Mayank Kejriwal

Although multiple COVID-19 vaccines have been available for several months now, vaccine hesitancy continues to be at high levels in the United States. In part, the issue has also become politicized, especially since the presidential election in November. Understanding vaccine hesitancy during this period in the context of social media, including Twitter, can provide valuable guidance both to computational social scientists and policy makers. Rather than studying a single Twitter corpus, this paper takes a novel view of the problem by comparatively studying two Twitter datasets collected between two different time periods (one before the election, and the other, a few months after) using the same, carefully controlled data collection and filtering methodology. Our results show that there was a significant shift in discussion from politics to COVID-19 vaccines from fall of 2020 to spring of 2021. By using clustering and machine learning-based methods in conjunction with sampling and qualitative analysis, we uncover several fine-grained reasons for vaccine hesitancy, some of which have become more (or less) important over time. Our results also underscore the intense polarization and politicization of this issue over the last year.

Trevor Bonjour, Marina Haliem, Aala Alsalem et al.

Learning to adapt and make real-time informed decisions in a dynamic and complex environment is a challenging problem. Monopoly is a popular strategic board game that requires players to make multiple decisions during the game. Decision-making in Monopoly involves many real-world elements such as strategizing, luck, and modeling of opponent's policies. In this paper, we present novel representations for the state and action space for the full version of Monopoly and define an improved reward function. Using these, we show that our deep reinforcement learning agent can learn winning strategies for Monopoly against different fixed-policy agents. In Monopoly, players can take multiple actions even if it is not their turn to roll the dice. Some of these actions occur more frequently than others, resulting in a skewed distribution that adversely affects the performance of the learning agent. To tackle the non-uniform distribution of actions, we propose a hybrid approach that combines deep reinforcement learning (for frequent but complex decisions) with a fixed policy approach (for infrequent but straightforward decisions). Experimental results show that our hybrid agent outperforms a standard deep reinforcement learning agent by 30% in the number of games won against fixed-policy agents.

Ke Shen, Mayank Kejriwal

Acquiring commonsense knowledge and reasoning is recognized as an important frontier in achieving general Artificial Intelligence (AI). Recent research in the Natural Language Processing (NLP) community has demonstrated significant progress in this problem setting. Despite this progress, which is mainly on multiple-choice question answering tasks in limited settings, there is still a lack of understanding (especially at scale) of the nature of commonsense knowledge itself. In this paper, we propose and conduct a systematic study to enable a deeper understanding of commonsense knowledge by doing an empirical and structural analysis of the ConceptNet knowledge base. ConceptNet is a freely available knowledge base containing millions of commonsense assertions presented in natural language. Detailed experimental results on three carefully designed research questions, using state-of-the-art unsupervised graph representation learning ('embedding') and clustering techniques, reveal deep substructures in ConceptNet relations, allowing us to make data-driven and computational claims about the meaning of phenomena such as 'context' that are traditionally discussed only in qualitative terms. Furthermore, our methodology provides a case study in how to use data-science and computational methodologies for understanding the nature of an everyday (yet complex) psychological phenomenon that is an essential feature of human intelligence.

Mayank Kejriwal, Ke Shen

Recently, transformer-based methods such as RoBERTa and GPT-3 have led to significant experimental advances in natural language processing tasks such as question answering and commonsense reasoning. The latter is typically evaluated through multiple benchmarks framed as multiple-choice instances of the former. According to influential leaderboards hosted by the Allen Institute (evaluating state-of-the-art performance on commonsense reasoning benchmarks), models based on such transformer methods are approaching human-like performance and have average accuracy well over 80% on many benchmarks. Since these are commonsense benchmarks, a model that generalizes on commonsense reasoning should not experience much performance loss across multiple commonsense benchmarks. In this paper, we study the generalization issue in detail by designing and conducting a rigorous scientific study. Using five common benchmarks, multiple controls and statistical analysis, we find clear evidence that fine-tuned commonsense language models still do not generalize well, even with moderate changes to the experimental setup, and may, in fact, be susceptible to dataset bias. We also perform selective studies, including qualitative and consistency analyses, to gain deeper insight into the problem.

Ravi Kiran Selvam, Mayank Kejriwal

Schema.org has experienced high growth in recent years. Structured descriptions of products embedded in HTML pages are now not uncommon, especially on e-commerce websites. The Web Data Commons (WDC) project has extracted schema.org data at scale from webpages in the Common Crawl and made it available as an RDF `knowledge graph' at scale. The portion of this data that specifically describes products offers a golden opportunity for researchers and small companies to leverage it for analytics and downstream applications. Yet, because of the broad and expansive scope of this data, it is not evident whether the data is usable in its raw form. In this paper, we do a detailed empirical study on the product-specific schema.org data made available by WDC. Rather than simple analysis, the goal of our study is to devise an empirically grounded set of best practices for using and consuming WDC product-specific schema.org data. Our studies reveal six best practices, each of which is justified by experimental data and analysis.

Jiayuan Ding, Mayank Kejriwal

Emotion Cause Extraction (ECE) aims to identify emotion causes from a document after annotating the emotion keywords. Some baselines have been proposed to address this problem, such as rule-based, commonsense based and machine learning methods. We show, however, that a simple random selection approach toward ECE that does not require observing the text achieves similar performance compared to the baselines. We utilized only position information relative to the emotion cause to accomplish this goal. Since position information alone without observing the text resulted in higher F-measure, we therefore uncovered a bias in the ECE single genre Sina-news benchmark. Further analysis showed that an imbalance of emotional cause location exists in the benchmark, with a majority of cause clauses immediately preceding the central emotion clause. We examine the bias from a linguistic perspective, and show that high accuracy rate of current state-of-art deep learning models that utilize location information is only evident in datasets that contain such position biases. The accuracy drastically reduced when a dataset with balanced location distribution is introduced. We therefore conclude that it is the innate bias in this benchmark that caused high accuracy rate of these deep learning models in ECE. We hope that the case study in this paper presents both a cautionary lesson, as well as a template for further studies, in interpreting the superior fit of deep learning models without checking for bias.

Mayank Kejriwal, Peilin Zhou

Humanitarian disasters have been on the rise in recent years due to the effects of climate change and socio-political situations such as the refugee crisis. Technology can be used to best mobilize resources such as food and water in the event of a natural disaster, by semi-automatically flagging tweets and short messages as indicating an urgent need. The problem is challenging not just because of the sparseness of data in the immediate aftermath of a disaster, but because of the varying characteristics of disasters in developing countries (making it difficult to train just one system) and the noise and quirks in social media. In this paper, we present a robust, low-supervision social media urgency system that adapts to arbitrary crises by leveraging both labeled and unlabeled data in an ensemble setting. The system is also able to adapt to new crises where an unlabeled background corpus may not be available yet by utilizing a simple and effective transfer learning methodology. Experimentally, our transfer learning and low-supervision approaches are found to outperform viable baselines with high significance on myriad disaster datasets.

Yao Gu, Mayank Kejriwal

In social media like Twitter, hashtags carry a lot of semantic information and can be easily distinguished from the main text. Exploring and visualizing the space of hashtags in a meaningful way can offer important insights into a dataset, especially in crisis situations. In this demonstration paper, we present a functioning prototype, HashViz, that ingests a corpus of tweets collected in the aftermath of a crisis situation (such as the Las Vegas shootings) and uses the fastText bag-of-tricks semantic embedding algorithm (from Facebook Research) to embed words and hashtags into a vector space. Hashtag vectors obtained in this way can be visualized using the t-SNE dimensionality reduction algorithm in 2D. Although multiple Twitter visualization platforms exist, HashViz is distinguished by being simple, scalable, interactive and portable enough to be deployed on a server for million-tweet corpora collected in the aftermath of arbitrary disasters, without special-purpose installation, technical expertise, manual supervision or costly software or infrastructure investment. Although simple, we show that HashViz offers an intuitive way to summarize, and gain insight into, a developing crisis situation. HashViz is also completely unsupervised, requiring no manual inputs to go from a raw corpus to a visualization and search interface. Using the recent Las Vegas mass shooting massacre as a case study, we illustrate the potential of HashViz using only a web browser on the client side.

Mayank Kejriwal, Yao Gu

Due to instant availability of data on social media platforms like Twitter, and advances in machine learning and data management technology, real-time crisis informatics has emerged as a prolific research area in the last decade. Although several benchmarks are now available, especially on portals like CrisisLex, an important, practical problem that has not been addressed thus far is the rapid acquisition and benchmarking of data from free, publicly available streams like the Twitter API. In this paper, we present ongoing work on a pipeline for facilitating immediate post-crisis data collection, curation and relevance filtering from the Twitter API. The pipeline is minimally supervised, alleviating the need for feature engineering by including a judicious mix of data preprocessing and fast text embeddings, along with an active learning framework. We illustrate the utility of the pipeline by describing a recent case study wherein it was used to collect and analyze millions of tweets in the immediate aftermath of the Las Vegas shootings.

Kyle Hundman, Thamme Gowda, Mayank Kejriwal et al.

Web-based human trafficking activity has increased in recent years but it remains sparsely dispersed among escort advertisements and difficult to identify due to its often-latent nature. The use of intelligent systems to detect trafficking can thus have a direct impact on investigative resource allocation and decision-making, and, more broadly, help curb a widespread social problem. Trafficking detection involves assigning a normalized score to a set of escort advertisements crawled from the Web -- a higher score indicates a greater risk of trafficking-related (involuntary) activities. In this paper, we define and study the problem of trafficking detection and present a trafficking detection pipeline architecture developed over three years of research within the DARPA Memex program. Drawing on multi-institutional data, systems, and experiences collected during this time, we also conduct post hoc bias analyses and present a bias mitigation plan. Our findings show that, while automatic trafficking detection is an important application of AI for social good, it also provides cautionary lessons for deploying predictive machine learning algorithms without appropriate de-biasing. This ultimately led to integration of an interpretable solution into a search system that contains over 100 million advertisements and is used by over 200 law enforcement agencies to investigate leads.

Rahul Kapoor, Mayank Kejriwal, Pedro Szekely

Extracting geographical tags from webpages is a well-motivated application in many domains. In illicit domains with unusual language models, like human trafficking, extracting geotags with both high precision and recall is a challenging problem. In this paper, we describe a geotag extraction framework in which context, constraints and the openly available Geonames knowledge base work in tandem in an Integer Linear Programming (ILP) model to achieve good performance. In preliminary empirical investigations, the framework improves precision by 28.57% and F-measure by 36.9% on a difficult human trafficking geotagging task compared to a machine learning-based baseline. The method is already being integrated into an existing knowledge base construction system widely used by US law enforcement agencies to combat human trafficking.

Mayank Kejriwal, Pedro Szekely

We propose a supervised algorithm for generating type embeddings in the same semantic vector space as a given set of entity embeddings. The algorithm is agnostic to the derivation of the underlying entity embeddings. It does not require any manual feature engineering, generalizes well to hundreds of types and achieves near-linear scaling on Big Graphs containing many millions of triples and instances by virtue of an incremental execution. We demonstrate the utility of the embeddings on a type recommendation task, outperforming a non-parametric feature-agnostic baseline while achieving 15x speedup and near-constant memory usage on a full partition of DBpedia. Using state-of-the-art visualization, we illustrate the agreement of our extensionally derived DBpedia type embeddings with the manually curated domain ontology. Finally, we use the embeddings to probabilistically cluster about 4 million DBpedia instances into 415 types in the DBpedia ontology.

Mayank Kejriwal, Pedro Szekely

Extracting useful entities and attribute values from illicit domains such as human trafficking is a challenging problem with the potential for widespread social impact. Such domains employ atypical language models, have `long tails' and suffer from the problem of concept drift. In this paper, we propose a lightweight, feature-agnostic Information Extraction (IE) paradigm specifically designed for such domains. Our approach uses raw, unlabeled text from an initial corpus, and a few (12-120) seed annotations per domain-specific attribute, to learn robust IE models for unobserved pages and websites. Empirically, we demonstrate that our approach can outperform feature-centric Conditional Random Field baselines by over 18\% F-Measure on five annotated sets of real-world human trafficking datasets in both low-supervision and high-supervision settings. We also show that our approach is demonstrably robust to concept drift, and can be efficiently bootstrapped even in a serial computing environment.

Janani Balaji, Faizan Javed, Mayank Kejriwal et al.

Entity Resolution, also called record linkage or deduplication, refers to the process of identifying and merging duplicate versions of the same entity into a unified representation. The standard practice is to use a Rule based or Machine Learning based model that compares entity pairs and assigns a score to represent the pairs' Match/Non-Match status. However, performing an exhaustive pair-wise comparison on all pairs of records leads to quadratic matcher complexity and hence a Blocking step is performed before the Matching to group similar entities into smaller blocks that the matcher can then examine exhaustively. Several blocking schemes have been developed to efficiently and effectively block the input dataset into manageable groups. At CareerBuilder (CB), we perform deduplication on massive datasets of people profiles collected from disparate sources with varying informational content. We observed that, employing a single blocking technique did not cover the base for all possible scenarios due to the multi-faceted nature of our data sources. In this paper, we describe our ensemble approach to blocking that combines two different blocking techniques to leverage their respective strengths.

Mayank Kejriwal

Several `edge-discovery' applications over graph-based data models are known to have worst-case quadratic time complexity in the nodes, even if the discovered edges are sparse. One example is the generic link discovery problem between two graphs, which has invited research interest in several communities. Specific versions of this problem include link prediction in social networks, ontology alignment between metadata-rich RDF data, approximate joins, and entity resolution between instance-rich data. As large datasets continue to proliferate, reducing quadratic complexity to make the task practical is an important research problem. Within the entity resolution community, the problem is commonly referred to as blocking. A particular class of learnable blocking schemes is known as Disjunctive Normal Form (DNF) blocking schemes, and has emerged as state-of-the art for homogeneous (i.e. same-schema) tabular data. Despite the promise of these schemes, a formalism or learning framework has not been developed for them when input data instances are generic, attributed graphs possessing both node and edge heterogeneity. With such a development, the complexity-reducing scope of DNF schemes becomes applicable to a variety of problems, including entity resolution and type alignment between heterogeneous graphs, and link prediction in networks represented as attributed graphs. This paper presents a graph-theoretic formalism for DNF schemes, and investigates their learnability in an optimization framework. We also briefly describe an empirical case study encapsulating some of the principles in this paper.