arXiv subjects

Harsh Kumar

Publications and source records attributed to Harsh Kumar.

At least 19 recordsLinked to original sources

arXiv2026

AT 2022csn: A Photometrically Peculiar Optical/UV Tidal Disruption Event in a Type II AGN

The emission mechanism and host galaxy preference of optical/UV tidal disruption events (TDEs) are still not entirely understood. We present observations of the TDE AT 2022csn, which is one of the most distant (d_L~726 Mpc) and luminous (L_peak=2.487^(+0.073)_(-0.067)*10^(44) erg/s) optical/UV TDEs observed to date. Although it is a spectroscopically normal H+He TDE, it shows some photometric peculiarities, exhibiting a pronounced double-peaked light curve (with peaks separated by 18.30 pm 2.84 days in the g-band), and lying in the low-temperature and large-radius end of the optical/UV TDE population. The host galaxy of AT 2022csn shows evidence for a significant starburst within the last ~Gyr consistent with other optical/UV TDEs, but also narrow emission lines that place it within the Type II AGN region of the BPT diagram. Interaction between the TDE and a pre-existing AGN accretion disk might explain the peculiar photometric properties. However, it is puzzling that a TDE would be visible in a Type II AGN, where according to the AGN unification picture the central region around the supermassive black hole is obscured. We suggest a few scenarios to reconcile this. AT 2022csn together with AT 2019ahk, which shows similar properties, may belong to a new subset of low-temperature, high-radius TDEs in galaxies with Type II AGN emission features.

astro-ph.HE↗

arXiv2026

Diagnosing and Repairing Persona Collapse in LLM Advice

LLMs are increasingly used for personal advice on relationships, work, moral dilemmas, and crises. Post-training selects a stable, prosocial Assistant persona, but good advice requires more than a good default character: a skilled advisor comforts someone in crisis, challenges someone in denial, and stays procedural with a logistical question. We formalize advice-giving as situation-conditioned persona selection in a space defined by hedonic tone and agency support, and call failures of this mapping "persona collapse" (the compression of diverse situations into a single default persona). Across 1,281 advice posts spanning 14 contexts, top-rated human responses shift systematically across five personas, while three frontier models collapse over 90\% of responses into a single supportive persona regardless of context. Prompting the model to first pick a fitting persona only deepens the collapse. We then ask whether the collapse can be repaired. Our method, Inverse-Process Distillation, reconstructs the situational reading that could have produced each human response and trains on the result, aiming to distill the situation-to-persona policy rather than the answers. It cuts divergence from the human persona distribution by approximately 80\%. Yet in a blinded study, 199 experienced advice-givers rating responses across four situations in sequence prefer the collapsed default over every repaired model, most strongly when the situation calls for challenge, though this preference shifts with repeated exposures.

cs.CY↗

arXiv2026

FedProIn: Mitigating Client Drift for Learnable Prototypes in Federated Medical Imaging

Federated learning (FL) is severely hindered by statistical heterogeneity due to variations in scanners, acquisition protocols, and patient populations. Such non-IID data induces client drift during local optimization, leading to unstable convergence and suboptimal global models when parameter-based aggregation is applied. We propose a prototype-based, influence-aware federated learning framework (FedProIn) that uses multiple learnable class prototypes to capture shared semantic structures across heterogeneous clients. We introduce feature divergence loss and prototype contrastive loss to mitigate client drift by decomposing it into feature drift and prototype drift. In addition, we propose a normalized influence aggregation strategy that adaptively weights client prototypes according to their contribution to the global representation, reducing the impact of biased or low-quality updates. Experimental results on two publicly available medical datasets, HAM10000 and Matek-19, demonstrate that FedProIn achieves accuracies of (83.5% IID, 81.1% non-IID) on HAM10000 and (96.2% IID, 95.8% non-IID) on Matek-19, respectively, outperforming existing baselines in both conditions. Our code is available at https://github.com/harsh-kmr/FedProIn.

eess.IV↗

arXiv2026

Position: We Need Large Language Models Optimized For Our Well-Being

Large language models are useful because we taught them to give us what we want. This works when success can be judged immediately, but people increasingly bring these systems their relationships, hard decisions, and long-term goals, where what a user wants to hear and what serves them best are frequently different. We argue that LLM providers should offer at least one widely accessible, opt-in mode optimized and evaluated for long-term well-being rather than next-turn approval. This is a pressing need, as models have been found to endorse questionable framings well above human baselines, users take AI advice readily without their well-being improving, and sycophantic models raise dependence while lowering prosocial intent. The mentors, coaches, and therapists we trust with our long-term development earn that trust by being willing to say what we do not want to hear, and LLMs should do the same. We propose three principles---change the objective, give users explicit relational roles, avoid paternalism---and organize the design space around three choices the current objective makes implicitly: the horizon over which well-being is measured (When), whose interests it represents (Who), and what role the assistant plays (How).

cs.CY↗

arXiv2026

The pair-instability origin of supernova 2023vbw

Stars in the initial and carbon-oxygen core mass ranges of $\sim140-260$ and $50-130$ M$_\odot$, respectively, with low metallicity are predicted to experience copious electron-positron pair production in their cores, leading to a runaway thermonuclear explosion that obliterates the entire star in a luminous and long-duration pair-instability supernova explosion. Some previous supernovae have been interpreted in this context but lack the full range of predicted properties. Here, we report detailed observations and modeling of the hydrogen-rich supernova 2023vbw, which exploded in a low-metallicity ($\sim0.1$ Z$_\odot$) environment in a dwarf star-forming galaxy at a redshift of $0.088$. Its light curve exhibits a luminous ($1.6\times10^{43}$ erg s$^{-1}$) and long-duration ($190$ days) main peak, resulting in a total radiated energy of $3\times10^{50}$ erg, more than an order of magnitude greater than canonical core-collapse supernovae. Semi-analytical light-curve modeling yields a blue supergiant-like progenitor with an ejecta mass of $170-350$ M$_\odot$, radioactive nickel mass of $1.2-1.6$ M$_\odot$, and explosion energy of $(6-13)\times10^{52}$ erg, well matched by pair-instability models. The early and late-phase light curve and spectra also show evidence for interaction of the supernova ejecta with an aspherical circumstellar medium. Discoveries of numerous such events with the upcoming Rubin Observatory and Roman Space Telescope will shed light on the deaths of the most massive stars in the Universe.

astro-ph.HE↗

arXiv2026

PermaFrost-Attack: Stealth Pretraining Seeding(SPS) for planting Logic Landmines During LLM Training

Aligned large language models (LLMs) remain vulnerable to adversarial manipulation, and their reliance on web-scale pretraining creates a subtle but consequential attack surface. We study Stealth Pretraining Seeding (SPS), a threat model in which adversaries distribute small amounts of poisoned content across stealth websites, increasing the likelihood that such material 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 that remains largely invisible under standard evaluation, yet can later be activated by a precise trigger such as <00TRIGGER00>. We call this attack PermaFrost, reflecting its latent and reactivatable nature. We study it through PermaFrost-Attack, a controlled framework for latent conceptual poisoning, together with three geometric diagnostics: Thermodynamic Length, Spectral Curvature, and the Infection Traceback Graph. Across multiple model families and scales, we show that this controlled SPS proxy can induce persistent unsafe behavior that often remains hidden under standard evaluation. 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 under standard evaluation.

cs.LG↗

arXiv2026

Human Thinking under Plural LLM Assistance: Mathematical Problem Solving and Open-Ended Writing

Large language models are changing not only the kind of assistance people receive, but also how that assistance is organized. Instead of working with a single general-purpose chatbot, people can now receive help from systems arranged as peers, specialists, or multiple agents with distinct roles. However, it remains unclear how these forms of plural LLM assistance affect human performance, confidence, and diversity of thought. We conducted two controlled experiments involving 562 participants to examine the effects of using multiple LLMs on mathematical problem-solving and writing. In a math task, participants worked with no LLM, an expert assistant, peer-like agents that surfaced common errors, or both an expert and a peer-like assistant. The expert-plus-peer condition produced the strongest unassisted post-task performance. In a writing task, participants wrote with no LLM, a single generalist assistant, or a pair of role-specialized assistants. LLM assistance improved essay quality, but the role-specialized pair preserved greater idea diversity than the single assistant. Together, these findings identify the arrangement of LLM assistance as a consequential design variable for human-AI collaboration.

cs.HC↗

arXiv2026

Investigating the Effects of LLM Use on Critical Thinking Under Time Constraints: Access Timing and Time Availability

The impact of large language models (LLMs) on critical thinking has provoked growing attention, yet this impact on actual performance may not be uniformly negative or positive. Particularly, the role of time -- the temporal context under which an LLM is provided -- remains overlooked. In a between-subjects experiment (n=393), we examined two types of time constraints for a critical thinking task requiring participants to make a reasoned decision for a real-world scenario based on diverse documents: (1) LLM access timing -- an LLM available only at the beginning (early), throughout (continuous), near the end (late), or not at all (no LLM), and (2) time availability -- insufficient or sufficient time for the task. We found a temporal reversal: LLM access from the start (early, continuous) improved performance under time pressure but impaired it with sufficient time, whereas beginning the task independently (late, no LLM) showed the opposite pattern. These findings demonstrate that time constraints fundamentally shape whether an LLM augments or undermines critical thinking, making time a central consideration when designing LLM support and evaluating human-AI collaboration in cognitive tasks.

cs.HC↗

arXiv2026

Transforming GenAI Policy to Prompting Instruction: An RCT of Scalable Prompting Interventions in a CS1 Course

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.

cs.HC↗

arXiv2026

Multiwavelength Analysis of Six Luminous Fast Blue Optical Transients

We present multiwavelength observations and analysis of six luminous fast blue optical transients (LFBOTs) discovered in Zwicky Transient Facility (ZTF) survey data. We identified these LFBOTs from their fast light-curve evolution ($t_{1/2}\leq 12 $d), blue colors at peak brightness ($g-r\leq-0.5 $mag), a visible host galaxy, high optical luminosity ($M_g<-20$), and an X-ray or radio detection. With the exception of AT2024aehp (ZTF24abygbss), these transients exhibit peaks in their $10\,$GHz radio light curves at $t_{\text{rest}} \approx 50-100$ d, with peak radio luminosities ranging from $10^{38}-10^{40}$ erg s$^{-1}$. Modeling the radio emission as synchrotron radiation indicates a fast ($v=0.1-0.3c$) shock in a dense ($n_e\approx10^{3}-10^{4}$ cm$^{-3}$) medium. The X-ray emission varies by $\approx2$ orders of magnitude in luminosity ($10^{42}-10^{44}$ erg s$^{-1}$) at $t_{\text{rest}}\sim20 $d. Analysis of the host-galaxy photometry and spectroscopy for each transient shows that they are predominantly nonnuclear (a few kpc offset) with star-forming host galaxies of stellar masses $10^{9}-10^{11} ,M_\odot$. Unlike all other LFBOTs to date, AT2024aehp exhibited a luminous ($M<-19 $mag) plateau in the optical light curve; spectra during this plateau phase showed a featureless blue continuum. The $6-15$ GHz radio emission of AT2024aehp brightened by over an order of magnitude from $t_{\text{rest}} \approx70 $d to $t_{\mathrm{rest}} \approx130 $d. The mostly consistent radio behavior between optically selected LFBOTs implies a similar circumburst medium, leading us to prefer a progenitor scenario in which mass is lost in a consistent way shortly prior to the terminal event, such as a massive star merging with a compact object.

astro-ph.HE↗

arXiv2025

SN 2024afav: A Superluminous Supernova with Multiple Light Curve Bumps and Spectroscopic Signatures of Circumstellar Interaction

We present a comprehensive optical and near-infrared spectroscopic study of SN 2024afav - a hydrogen-poor superluminous supernova (SLSN-I) that peaks at $\approx$ -20.7 mag and exhibits an unusual multi-bumped light curve. Our spectroscopic observations, spanning phases of -14 to +160 d, reveal several unusual features: (i) a narrow (1,800 km s$^{-1}$) and blueshifted (11,000 km s$^{-1}$) absorption from H$\alpha$ starting at +20 d; (ii) persistent optical and NIR He I lines at all available phases, showing double absorption structure in NIR spectra at +23 d, with a high velocity component at a similar velocity to H$\alpha$; (iii) early appearance of nebular [O III] emission starting at $\approx$ +50 d; and (iv) strong [O II] + [Ca II] 7300 {\AA} emission complex starting at $\approx$ +110 d. These unusual features, and their onset at the time of the light curve bumps, provide compelling evidence of circumstellar interaction between the SN ejecta and a nearby hydrogen-rich shell, as well as the presence of helium in both the outer layers of the progenitor star and in the circumstellar medium. A comparison of SN 2024afav to other SLSNe-I showing bumpy light curves and similar spectral properties (PTF10hgi, SN 2017egm, SN 2019hge), points to a rare sub-group of SLSNe-I in which CSM interaction provides an important modulation to the energy input.

astro-ph.SR↗

arXiv2025

GRB 230204B: GIT Discovery of a Fast Fading Afterglow Associated with an Energetic GRB from a Massive-Star Progenitor

We present a comprehensive multi-wavelength study of a bright gamma-ray burst GRB 230204B, analyzing both prompt and afterglow emissions. This GRB is highly energetic, with an isotropic equivalent energy emission $E_{\mathrm{iso}} \sim 2.2 \times 10^{54}\ \mathrm{erg}$, released during the prompt emission. The GROWTH-India Telescope discovered a bright afterglow ($m_r = 15.55$) that faded rapidly ($\propto t^{-1.82}$). The prompt emission shows strong thermal photospheric emission, along with a non-thermal high-energy component. We explore the evolution of these components and find them to be consistent with theoretical expectations. Afterglow modeling reveals an energetic jet $E_{tot} \gtrsim 10^{52}\ \mathrm{erg}$ expanding into a wind-type medium viewed nearly on-axis, suggesting a massive star progenitor with strong winds. We also explore correlations between the prompt emission and afterglow that may help to understand the complete picture of GRB progenitors.

astro-ph.HE↗

arXiv2025

"When Data is Scarce, Prompt Smarter"... Approaches to Grammatical Error Correction in Low-Resource Settings

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.

cs.CL↗

arXiv2025

Hydrogen-Poor Superluminous Supernovae in the Nebular Phase: Spectral Diversity Due to Ejecta Ionization as a Probe of the Power Source

We present a large sample of 39 nebular-phase optical spectra of 25 hydrogen-poor superluminous supernovae (SLSNe-I) and jointly analyze them with previously published spectra of 12 events. We measure the properties of key emission features, namely those at 6300, 7300, and 7774 angstroms (associated with [O I], [Ca II]/[O II], and O I, respectively), and find that SLSNe exhibit much wider spectral diversity than normal SNe Ic, primarily in the line ratio $L_{7300}/L_{6300}$, which is highly sensitive to ejecta ionization. Some events exhibit weak [O I] and a clear [O II] contribution to the 7300 angstrom feature, enhancing the ratio, along with [O III] lines at 4363 and 5007 angstroms. Other SLSNe show weak or no lines of ionized oxygen. Moreover, we find that the population exhibits decreasing $L_{7300}/L_{6300}$ over time, while a few outliers instead display sustained high or increasing ratios for extended periods. The ratio $L_{7300}/L_{6300}$ is also correlated with the rise and decline times of the light curves, with slower events exhibiting higher ionization, the first robust connection between early light curve and late-time spectral properties, likely due to the magnetar's impact: slower-evolving SLSNe are generally powered by engines with longer spin-down timescales, which deposit more energy at later phases. Among the events with decreasing $L_{7300}/L_{6300}$, SLSNe with high ionization are on average powered by magnetars with higher thermalized spin-down power, a correlation that is most significant for events with $M_{\rm ej}\lesssim12$ M$_{\odot}$. The ionization in the outliers with increasing $L_{7300}/L_{6300}$ may be due to late CSM interaction. $L_{7300}/L_{6300}$ and its evolution are therefore key diagnostics of SLSN engines and progenitor mass loss.

astro-ph.HE↗

arXiv2025

When AI Gives Advice: Evaluating AI and Human Responses to Online Advice-Seeking for Well-Being

Seeking advice is a core human behavior that the internet has reinvented twice: first through forums and Q&A communities that crowdsource public guidance, and now through large language models (LLMs). Yet the quality of this LLM advice for everyday well-being scenarios remains unclear. How does it compare, not only against human comments, but against the wisdom of the online crowd? We ran two studies (N=210) in which experts compared top-voted Reddit advice with LLM-generated advice. LLMs ranked significantly higher overall and on effectiveness, warmth, and willingness to seek advice again. GPT-4o beat GPT-5 on all metrics except sycophancy, suggesting that benchmark gains need not improve advice-giving. In Study-2, we examined how human and algorithmic advice could be combined, and found that human advice can be unobtrusively polished to compete with AI-generated comments. We conclude with design implications for advice-giving agents and ecosystems blending AI, crowd input, and expert oversight.

cs.HC↗

arXiv2025

Optimizing Kilonova Searches: A Case Study of the Type IIb SN 2025ulz in the Localization Volume of the Low-Significance Gravitational Wave Event S250818k

Kilonovae, the ultraviolet/optical/infrared counterparts to binary neutron star mergers, are an exceptionally rare class of transients. Optical follow-up campaigns are plagued by contaminating transients, which may mimic kilonovae, but do not receive sufficient observations to measure the full photometric evolution. In this work, we present an analysis of the multi-wavelength dataset of supernova (SN) 2025ulz, a proposed kilonova candidate following the low-significance detection of gravitational waves originating from the potential binary neutron star merger S250818k. Despite an early rapid decline in brightness, our multi-wavelength observations of SN 2025ulz reveal that it is a type IIb supernova. As part of this analysis, we demonstrate the capabilities of a novel quantitative scoring algorithm to determine the likelihood that a transient candidate is a kilonova, based primarily on its 3D location and light curve evolution. We also apply our scoring algorithm to other transient candidates in the localization volume of S250818k and find that, at all times after the discovery of SN 2025ulz, there are $\geq 4$ candidates with a score comparable to SN 2025ulz, indicating that the kilonova search may have benefited from the additional follow-up of other candidates. During future kilonova searches, this type of scoring algorithm will be useful to rule out contaminating transients in real time, optimizing the use of valuable telescope resources.

astro-ph.HE↗

arXiv2025

When IIb Ceases To Be: Bridging the Gap Between IIb and Short-plateau Supernovae

Hydrogen-rich supernovae (SNe) span a range of hydrogen envelope masses at core collapse, producing diverse light curves from extended plateaus in Type II SNe to double-peaked Type IIb SNe. Recent hydrodynamic modeling predicts a continuous sequence of light-curve morphologies as hydrogen is removed, with short plateau SNe (plateau durations ~50--70 days) emerging as a transitional class. However, the observational boundary between IIb and short-plateau remains poorly defined, and thus far unobserved. We report on extensive photometric and spectroscopic follow-up of SN 2023wdd and SN 2022acrv, candidate transitional events on the low-mass end of the short-plateau class. Both exhibit weak, double-peaked light curves which we interpret as exceptionally short plateaus (10--20 days), and hybrid spectral features: persistent H$\alpha$ absorption with He I contamination, but without the helium dominance characteristic of IIb SNe. Using analytic shock-cooling models and numerical light curve fitting, we estimate hydrogen-rich envelope masses of ~0.6--0.8 $M_\odot$ -- significantly larger than canonical IIb values ($\lesssim0.1\,M_\odot$) but consistent with the ${\sim}0.9\,M_\odot$ threshold predicted for short-plateau behavior. Although the progenitor radii inferred from analytic and numerical methods differ by factors of 2--5, envelope mass estimates are consistent across approaches. Comparisons to well-studied IIb (SN 2016gkg, SN 2022hnt), short-plateau (SN 2023ufx, SN 2006ai, SN 2016egz, SN 2006Y), and II SNe (SN 2023ixf, SN 2013ej) suggest a monotonic relationship between hydrogen envelope mass and plateau length consistent with analytic and numerical expectations. These findings provide additional evidence for a continuous distribution of envelope stripping in hydrogen-rich core-collapse progenitors and place SN 2023wdd and SN 2022acrv along the IIb/short-plateau boundary.

astro-ph.HE↗

arXiv2025

Lense-Thirring precessing magnetar engine drives a superluminous supernova

Type I superluminous supernovae (SLSNe-I) are at least an order of magnitude brighter than standard supernovae, with the internal power source for their luminosity still unknown. The central engines of SLSNe-I are hypothesized to be magnetars, but the majority of SLSNe-I light curves have multiple bumps or peaks that are unexplained by the standard magnetar model. Existing explanations for the bumps either modulate the central engine luminosity or invoke interactions with material in the circumstellar environment. Systematic surveys of the limited sample of SLSNe-I light curves find no compelling evidence favoring either scenario, leaving both the nature of the light-curve fluctuations and the applicability of the magnetar model unresolved. Here, we report high-cadence multiband observations of an SLSN-I with clear "chirped" (i.e., decreasing period) light-curve bumps that can be directly linked to the properties of the magnetar central engine. Our observations are consistent with a tilted, infalling accretion disk undergoing Lense-Thirring precession around a magnetar centrally located within the expanding supernova ejecta. Our model demonstrates that the overall light curve and bump frequency independently and self-consistently constrain the spin period and the magnetic field strength of the magnetar. Assuming standard accretion disk parameters, we constrain the accretion rate onto the magnetar. Our results provide the first observational evidence of the Lense-Thirring effect in the environment of a magnetar, and confirm the magnetar spin-down model as an explanation for the extreme luminosity observed in SLSNe-I. We anticipate this discovery will create avenues for testing general relativity in a new regime -- the violent centers of young supernovae.

astro-ph.HE↗