Harsh Kumar

Harsh Kumar, Ilya Musabirov, Mohi Reza et al.

Personalized chatbot-based teaching assistants can be crucial in addressing increasing classroom sizes, especially where direct teacher presence is limited. Large language models (LLMs) offer a promising avenue, with increasing research exploring their educational utility. However, the challenge lies not only in establishing the efficacy of LLMs but also in discerning the nuances of interaction between learners and these models, which impact learners' engagement and results. We conducted a formative study in an undergraduate computer science classroom (N=145) and a controlled experiment on Prolific (N=356) to explore the impact of four pedagogically informed guidance strategies on the learners' performance, confidence and trust in LLMs. Direct LLM answers marginally improved performance, while refining student solutions fostered trust. Structured guidance reduced random queries as well as instances of students copy-pasting assignment questions to the LLM. Our work highlights the role that teachers can play in shaping LLM-supported learning environments.

Harsh Kumar, Tong Li, Jiakai Shi et al.

Digital mental health (DMH) interventions, such as text-message-based lessons and activities, offer immense potential for accessible mental health support. While these interventions can be effective, real-world experimental testing can further enhance their design and impact. Adaptive experimentation, utilizing algorithms like Thompson Sampling for (contextual) multi-armed bandit (MAB) problems, can lead to continuous improvement and personalization. However, it remains unclear when these algorithms can simultaneously increase user experience rewards and facilitate appropriate data collection for social-behavioral scientists to analyze with sufficient statistical confidence. Although a growing body of research addresses the practical and statistical aspects of MAB and other adaptive algorithms, further exploration is needed to assess their impact across diverse real-world contexts. This paper presents a software system developed over two years that allows text-messaging intervention components to be adapted using bandit and other algorithms while collecting data for side-by-side comparison with traditional uniform random non-adaptive experiments. We evaluate the system by deploying a text-message-based DMH intervention to 1100 users, recruited through a large mental health non-profit organization, and share the path forward for deploying this system at scale. This system not only enables applications in mental health but could also serve as a model testbed for adaptive experimentation algorithms in other domains.

Harsh Kumar, Rahul Maity, Tanmay Joshi et al.

Aligned large language models(LLMs) remain vulnerable to adversarial manipulation, and their dependence on web-scale pretraining creates a subtle but serious attack surface. We study Stealth Pretraining Seeding (SPS), a new attack family in which adversaries distribute small amounts of poisoned content across stealth websites, expose them to web crawlers through robots.txt, and thereby increase the likelihood that such content is absorbed into future training corpora derived from sources such as Common Crawl. Because each individual payload is tiny, diffuse, and superficially benign, the attack is difficult to detect during dataset construction or filtering. The result is a latent form of poisoning: dormant logic landmines embedded during pretraining that remain largely invisible under standard evaluation, yet can later be activated by precise alphanumeric triggers such as <00TRIGGER00> to bypass safeguards. We call this attack PermaFrost, by analogy to Arctic permafrost: harmful material can remain frozen, buried, and unnoticed for long periods, only to resurface when conditions allow. We operationalize this threat through PermaFrost-Attack, a controlled framework for latent conceptual poisoning, together with a suite of geometric diagnostics: Thermodynamic Length, Spectral Curvature, and the Infection Traceback Graph. Across multiple model families and scales, we show that SPS is broadly effective, inducing persistent unsafe behavior while often evading alignment defenses. Our results identify SPS as a practical and underappreciated threat to future foundation models. This paper introduces a novel geometric diagnostic lens for systematically examining latent model behavior, providing a principled foundation for detecting, characterizing, and understanding vulnerabilities that may remain invisible to standard evaluation.

Harsh Kumar, Mohi Reza, Jeb Mitchell et al.

Growth in the use of large language models (LLMs) in programming education is altering how students write SQL queries. Traditionally, students relied heavily on web search for coding assistance, but this has shifted with the adoption of LLMs like ChatGPT. However, the comparative process and outcomes of using web search versus LLMs for coding help remain underexplored. To address this, we conducted a randomized interview study in a database classroom to compare web search and LLMs, including a publicly available LLM (ChatGPT) and an instructor-tuned LLM, for writing SQL queries. Our findings indicate that using an instructor-tuned LLM required significantly more interactions than both ChatGPT and web search, but resulted in a similar number of edits to the final SQL query. No significant differences were found in the quality of the final SQL queries between conditions, although the LLM conditions directionally showed higher query quality. Furthermore, students using instructor-tuned LLM reported a lower mental demand. These results have implications for learning and productivity in programming education.

Harsh Kumar, Suhyeon Yoo, Angela Zavaleta Bernuy et al.

Although engagement in self-directed wellness exercises typically declines over time, integrating social support such as coaching can sustain it. However, traditional forms of support are often inaccessible due to the high costs and complex coordination. Large Language Models (LLMs) show promise in providing human-like dialogues that could emulate social support. Yet, in-depth, in situ investigations of LLMs to support behavior change remain underexplored. We conducted two randomized experiments to assess the impact of LLM agents on user engagement with mindfulness exercises. First, a single-session study, involved 502 crowdworkers; second, a three-week study, included 54 participants. We explored two types of LLM agents: one providing information and another facilitating self-reflection. Both agents enhanced users' intentions to practice mindfulness. However, only the information-providing LLM, featuring a friendly persona, significantly improved engagement with the exercises. Our findings suggest that specific LLM agents may bridge the social support gap in digital health interventions.

Ruiwei Xiao, Runlong Ye, Xinying Hou et al. · utoronto

Despite universal GenAI adoption, students cannot distinguish task performance from actual learning and lack skills to leverage AI for learning, leading to worse exam performance when AI use remains unreflective. Yet few interventions teaching students to prompt AI as a tutor rather than solution provider have been validated at scale through randomized controlled trials (RCTs). To bridge this gap, we conducted a semester-long RCT (N=979) with four ICAP framework-based instructional conditions varying in engagement intensity with a pre-test, immediate and delayed post-test and surveys. Mixed methods analysis results showed: (1) All conditions significantly improved prompting skills, with gains increasing progressively from Condition 1 to Condition 4, validating ICAP's cognitive engagement hierarchy; (2) for students with similar pre-test scores, higher learning gain in immediate post-test predict higher final exam score, though no direct between-group differences emerged; (3) Our interventions are suitable and scalable solutions for diverse educational contexts, resources and learners. Together, this study makes empirical and theoretical contributions: (1) theoretically, we provided one of the first large-scale RCTs examining how cognitive engagement shapes learning in prompting literacy and clarifying the relationship between learning-oriented prompting skills and broader academic performance; (2) empirically, we offered timely design guidance for transforming GenAI classroom policies into scalable, actionable prompting literacy instruction to advance learning in the era of Generative AI.

Harsh Kumar, Zi Kang, Mu et al.

Most AI-based educational tools today adopt a one-on-one tutoring paradigm, pairing a single LLM with a single learner. Yet decades of learning science research suggest that multi-party interaction -- through peer modeling, co-construction, and exposure to diverse perspectives -- can produce learning benefits that dyadic tutoring alone cannot. In this paper, we investigate whether multi-agent LLM configurations can enhance learning outcomes beyond what a single LLM tutor provides. We present two controlled experiments spanning distinct learning contexts. In a convergent problem-solving study ($N=315$), participants tackle SAT-level math problems in a 2$\times$2 design that varies the presence of an LLM tutor and LLM peers, each making different kinds of errors (conceptual vs.\ arithmetic); participants who interacted with both a tutor and peers achieved the highest unassisted test accuracy. In a divergent composition study ($N=247$), participants write argumentative and creative essays with either no AI assistance, a single LLM (Claude or ChatGPT), or both Claude and ChatGPT together; while both LLM conditions improved essay quality, only the two-agent condition avoided the idea-level homogeneity that single-model assistance was found to produce. Together, these studies offer one of the first controlled investigations of multi-agent LLM learning environments, probing whether the move from one-on-one AI tutoring toward richer agent configurations can unlock the collaborative and observational benefits long documented in human social learning research.

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Somsubhra De, Harsh Kumar, Arun Prakash A

Grammatical error correction (GEC) is an important task in Natural Language Processing that aims to automatically detect and correct grammatical mistakes in text. While recent advances in transformer-based models and large annotated datasets have greatly improved GEC performance for high-resource languages such as English, the progress has not extended equally. For most Indic languages, GEC remains a challenging task due to limited resources, linguistic diversity and complex morphology. In this work, we explore prompting-based approaches using state-of-the-art large language models (LLMs), such as GPT-4.1, Gemini-2.5 and LLaMA-4, combined with few-shot strategy to adapt them to low-resource settings. We observe that even basic prompting strategies, such as zero-shot and few-shot approaches, enable these LLMs to substantially outperform fine-tuned Indic-language models like Sarvam-22B, thereby illustrating the exceptional multilingual generalization capabilities of contemporary LLMs for GEC. Our experiments show that carefully designed prompts and lightweight adaptation significantly enhance correction quality across multiple Indic languages. We achieved leading results in the shared task--ranking 1st in Tamil (GLEU: 91.57) and Hindi (GLEU: 85.69), 2nd in Telugu (GLEU: 85.22), 4th in Bangla (GLEU: 92.86), and 5th in Malayalam (GLEU: 92.97). These findings highlight the effectiveness of prompt-driven NLP techniques and underscore the potential of large-scale LLMs to bridge resource gaps in multilingual GEC.

Krish Goel, Sanskar Pandey, KS Mahadevan et al.

Human cognition is deeply intertwined with a sense of time, known as Chronoception. This sense allows us to judge how long facts remain valid and when knowledge becomes outdated. Despite progress in vision, language, and motor control, AI still struggles to reason about temporal validity. We introduce Chronocept, the first benchmark to model temporal validity as a continuous probability distribution over time. Using skew-normal curves fitted along semantically decomposed temporal axes, Chronocept captures nuanced patterns of emergence, decay, and peak relevance. It includes two datasets: Benchmark I (atomic facts) and Benchmark II (multi-sentence passages). Annotations show strong inter-annotator agreement (84% and 89%). Our baselines predict curve parameters - location, scale, and skewness - enabling interpretable, generalizable learning and outperforming classification-based approaches. Chronocept fills a foundational gap in AI's temporal reasoning, supporting applications in knowledge grounding, fact-checking, retrieval-augmented generation (RAG), and proactive agents. Code and data are publicly available.

Harsh Kumar

Many methods have been proposed to find vector representation for words, but most rely on capturing context from the text to find semantic relationships between these vectors. We propose a novel method of using dictionary meanings and image depictions to find word vectors independent of any context. We use auto-encoder on the word images to find meaningful representations and use them to calculate the word vectors. We finally evaluate our method on word similarity, concept categorization and outlier detection tasks. Our method performs comparably to context-based methods while taking much less training time.

Tong Li, Jacob Nogas, Haochen Song et al.

Traditional randomized A/B experiments assign arms with uniform random (UR) probability, such as 50/50 assignment to two versions of a website to discover whether one version engages users more. To more quickly and automatically use data to benefit users, multi-armed bandit algorithms such as Thompson Sampling (TS) have been advocated. While TS is interpretable and incorporates the randomization key to statistical inference, it can cause biased estimates and increase false positives and false negatives in detecting differences in arm means. We introduce a more Statistically Sensitive algorithm, TS-PostDiff (Posterior Probability of Small Difference), that mixes TS with traditional UR by using an additional adaptive step, where the probability of using UR (vs TS) is proportional to the posterior probability that the difference in arms is small. This allows an experimenter to define what counts as a small difference, below which a traditional UR experiment can obtain informative data for statistical inference at low cost, and above which using more TS to maximize user benefits is key. We evaluate TS-PostDiff against UR, TS, and two other TS variants designed to improve statistical inference. We consider results for the common two-armed experiment across a range of settings inspired by real-world applications. Our results provide insight into when and why TS-PostDiff or alternative approaches provide better tradeoffs between benefiting users (reward) and statistical inference (false positive rate and power). TS-PostDiff's adaptivity helps efficiently reduce false positives and increase statistical power when differences are small, while increasing reward more when differences are large. The work highlights important considerations for future Statistically Sensitive algorithm development that balances reward and statistical analysis in adaptive experimentation.

Harsh Kumar, Jithu Chandran

Shapley value has recently become a popular way to explain the predictions of complex and simple machine learning models. This paper is discusses the factors that influence Shapley value. In particular, we explore the relationship between the distribution of a feature and its Shapley value. We extend our analysis by discussing the difference that arises in Shapley explanation for different predicted outcomes from the same model. Our assessment is that Shapley value for particular feature not only depends on its expected mean but on other moments as well such as variance and there are disagreements for baseline prediction, disagreements for signs and most important feature for different outcomes such as probability, log odds, and binary decision generated using same linear probability model (logit/probit). These disagreements not only stay for local explainability but also affect the global feature importance. We conclude that there is no unique Shapley explanation for a given model. It varies with model outcome (Probability/Log-odds/binary decision such as accept vs reject) and hence model application.

Harsh Kumar, Geneva Allison, Jehil Mehta et al.

Well acquisition in the oil and gas industry can often be a hit or miss process, with a poor purchase resulting in substantial loss. Recommender systems suggest items (wells) that users (companies) are likely to buy based on past activity, and applying this system to well acquisition can increase company profits. While traditional recommender systems are impactful enough on their own, they are not optimized. This is because they ignore many of the complexities involved in human decision-making, and frequently make subpar recommendations. Using a preexisting Python implementation of a Factorization Machine results in more accurate recommendations based on a user-level ranking system. We train a Factorization Machine model on oil and gas well data that includes features such as elevation, total depth, and location. The model produces recommendations by using similarities between companies and wells, as well as their interactions. Our model has a hit rate of 0.680, reciprocal rank of 0.469, precision of 0.229, and recall of 0.463. These metrics imply that while our model is able to recommend the correct wells in a general sense, it does not match exact wells to companies via relevance. To improve the model's accuracy, future models should incorporate additional features such as the well's production data and ownership duration as these features will produce more accurate recommendations.