Search

We present CluStRE, a novel streaming graph clustering algorithm that balances computational efficiency with high-quality clustering using multi-stage refinement. Unlike traditional in-memory clustering approaches, CluStRE processes graphs in a streaming setting, significantly reducing memory overhead while leveraging re-streaming and evolutionary heuristics to improve solution quality. Our method dynamically constructs a quotient graph, enabling modularity-based optimization while efficiently handling large-scale graphs. We introduce multiple configurations of CluStRE to provide trade-offs between speed, memory consumption, and clustering quality. Experimental evaluations demonstrate that CluStRE improves solution quality by 89.8%, operates 2.6 times faster, and uses less than two-thirds of the memory required by the state-of-the-art streaming clustering algorithm on average. Moreover, our strongest mode enhances solution quality by up to 150% on average. With this, CluStRE achieves comparable solution quality to in-memory algorithms, i.e. over 96% of the quality of clustering approaches, including Louvain, effectively bridging the gap between streaming and traditional clustering methods.

AI Safety Moderation (ASM) classifiers are designed to moderate content on social media platforms and to serve as guardrails that prevent Large Language Models (LLMs) from being fine-tuned on unsafe inputs. Owing to their potential for disparate impact, it is crucial to ensure that these classifiers: (1) do not unfairly classify content belonging to users from minority groups as unsafe compared to those from majority groups and (2) that their behavior remains robust and consistent across similar inputs. In this work, we thus examine the fairness and robustness of four widely-used, closed-source ASM classifiers: OpenAI Moderation API, Perspective API, Google Cloud Natural Language (GCNL) API, and Clarifai API. We assess fairness using metrics such as demographic parity and conditional statistical parity, comparing their performance against ASM models and a fair-only baseline. Additionally, we analyze robustness by testing the classifiers' sensitivity to small and natural input perturbations. Our findings reveal potential fairness and robustness gaps, highlighting the need to mitigate these issues in future versions of these models., Comment: Accepted to NAACL 2025 Main Conference

Social media platforms utilize Machine Learning (ML) and Artificial Intelligence (AI) powered recommendation algorithms to maximize user engagement, which can result in inadvertent exposure to harmful content. Current moderation efforts, reliant on classifiers trained with extensive human-annotated data, struggle with scalability and adapting to new forms of harm. To address these challenges, we propose a novel re-ranking approach using Large Language Models (LLMs) in zero-shot and few-shot settings. Our method dynamically assesses and re-ranks content sequences, effectively mitigating harmful content exposure without requiring extensive labeled data. Alongside traditional ranking metrics, we also introduce two new metrics to evaluate the effectiveness of re-ranking in reducing exposure to harmful content. Through experiments on three datasets, three models and across three configurations, we demonstrate that our LLM-based approach significantly outperforms existing proprietary moderation approaches, offering a scalable and adaptable solution for harm mitigation., Comment: This paper is under peer review

The prevalence of harmful content on social media platforms poses significant risks to users and society, necessitating more effective and scalable content moderation strategies. Current approaches rely on human moderators, supervised classifiers, and large volumes of training data, and often struggle with scalability, subjectivity, and the dynamic nature of harmful content (e.g., violent content, dangerous challenge trends, etc.). To bridge these gaps, we utilize Large Language Models (LLMs) to undertake few-shot dynamic content moderation via in-context learning. Through extensive experiments on multiple LLMs, we demonstrate that our few-shot approaches can outperform existing proprietary baselines (Perspective and OpenAI Moderation) as well as prior state-of-the-art few-shot learning methods, in identifying harm. We also incorporate visual information (video thumbnails) and assess if different multimodal techniques improve model performance. Our results underscore the significant benefits of employing LLM based methods for scalable and dynamic harmful content moderation online., Comment: This paper is in submission and under peer review

This work introduces a novel paradigm for generalized In-Context Learning (ICL), termed Indirect In-Context Learning. In Indirect ICL, we explore demonstration selection strategies tailored for two distinct real-world scenarios: Mixture of Tasks and Noisy Demonstrations. We systematically evaluate the effectiveness of Influence Functions (IFs) as a selection tool for these settings, highlighting the potential for IFs to better capture the informativeness of examples within the demonstration pool. For the Mixture of Tasks setting, demonstrations are drawn from 28 diverse tasks, including MMLU, BigBench, StrategyQA, and CommonsenseQA. We demonstrate that combining BertScore-Recall (BSR) with an IF surrogate model can significantly improve performance, leading to average absolute accuracy gains of 0.37\% and 1.45\% for 3-shot and 5-shot setups when compared to traditional ICL metrics. In the Noisy Demonstrations setting, we examine scenarios where demonstrations might be mislabeled. Our experiments show that reweighting traditional ICL selectors (BSR and Cosine Similarity) with IF-based selectors boosts accuracy by an average of 2.90\% for Cosine Similarity and 2.94\% for BSR on noisy GLUE benchmarks. In sum, we propose a robust framework for demonstration selection that generalizes beyond traditional ICL, offering valuable insights into the role of IFs for Indirect ICL., Comment: Under Review

Many classical Be stars acquire their very rapid rotation by mass and angular-momentum transfer in massive binaries. Short-lived intermediate-phase objects have only been discovered recently. Data archives and the literature have been searched for additional candidates exhibiting this patterns. Thirteen candidates were identified at various confidence levels. Adding to the two known systems identified as classical Be star+pre-subdwarf binaries (LB-1 and HR6819), two more (V742Cas, HD44637) could be confirmed with interferometry, with V742Cas setting a new record for the smallest visually observed angular semi-major axis, at a=0.663mas. Two further ones (V447Sct, V1362Cyg) are not resolved interferometrically, but other evidence puts them at the same confidence level as LB-1. V2174Cyg is a candidate with very high confidence, but was not observed interferometrically. The remaining ones are either candidates with varying levels of confidence. Of a mostly magnitude complete sample of 328 Be stars, 0.5-1% are found to have recently completed the mass overflow that led to their formation. Another 5% are systems with compact subdwarf companions, i.e., further evolved after a previous overflow, and possibly two more percent harbor white dwarfs. All these systems are of early B-subtypes, however, and if the original sample is restricted to early subtypes (136 objects), these percentages increase by a factor of about 2.5, while dropping to zero for the mid and late subtypes (together 204 objects). This strongly suggests that early- vs. mid- and late-type Be stars have differently weighted channels to acquire their rapid rotation, namely binary interaction vs. evolutionary spin-up., Comment: Accepted for publication by A&A

The constant workspace algorithms use a constant number of words in addition to the read-only input to the algorithm stored in an array. In this paper, we devise algorithms to efficiently compute relative hulls in the plane using a constant workspace. Specifically, we devise algorithms for the following three problems: \newline (i) Given two simple polygons $P$ and $Q$ with $P \subseteq Q$, compute a simple polygon $P'$ with a perimeter of minimum length such that $P \subseteq P' \subseteq Q$. \newline (ii) Given two simple polygons $P$ and $Q$ such that $Q$ does not intersect the interior of $P$ but it does intersects with the interior of the convex hull of $P$, compute a weakly simple polygon $P'$ contained in the convex hull of $P$ such that the perimeter of $P'$ is of minimum length. \newline (iii) Given a set $S$ of points located in a simple polygon $P$, compute a weakly simple polygon $P' \subseteq P$ with a perimeter of minimum length such that $P'$ contains all the points in $S$. \newline To our knowledge, no prior works devised algorithms to compute relative hulls using a constant workspace and this work is the first such attempt.

The aim of this article is to give an expository account of the equivalence between modest sets and partial equivalence relations. Our proof is entirely self-contained in that we do not assume any knowledge of categorical realizability. At the heart of the equivalence lies the subquotient construction on a partial equivalence relation. The subquotient construction embeds the category of partial equivalence relations into the category of modest sets. We show that this embedding is a split essentially surjective functor, and thereby, an equivalence of categories. Our development is both constructive and predicative, and employs the language of homotopy type theory. All the mathematics presented in this article has been mechanised in Cubical Agda., Comment: 14 pages; Expository article

Cepheid masses are particularly necessary to help solving the mass discrepancy, while independent distance determinations provide crucial test of the period-luminosity relation and Gaia parallaxes. We used CHARA/MIRC to measure the astrometric positions of the high-contrast companion orbiting the Cepheid SU Cygni. We also present new radial velocity measurements from the HST. The combination of interferometric astrometry with optical and ultraviolet spectroscopy provides the full orbital elements of the system, in addition to component masses and the distance to the Cepheid system. We measured the mass of the Cepheid, $M_A = 4.859\pm0.058M_\odot$, and its two companions, $M_{Ba} = 3.595 \pm 0.033 M_\odot$ and $M_{Bb} = 1.546 \pm 0.009 M_\odot$. This is the most accurate existing measurement of the mass of a Galactic Cepheid (1.2%). Comparing with stellar evolution models, we show that the mass predicted is higher than the measured mass of the Cepheid, similar to conclusions of our previous work. We also measured the distance to the system to be $926.3 \pm 5.0$pc, i.e. an unprecedented parallax precision of $6\mu$as (0.5%), being the most precise and accurate distance for a Cepheid. Such precision is similar to what is expected by Gaia for the last data release (DR5 in $\sim$ 2030) for single stars fainter than G = 13, but is not guaranteed for stars as bright as SU Cyg. We demonstrated that evolutionary models remain inadequate in accurately reproducing the measured mass, often predicting higher masses for the expected metallicity, even when factors such as rotation or convective core overshooting are taken into account. Our precise distance measurement allowed us to compare prediction period-luminosity relations. We found a disagreement of 0.2-0.5 mag with relations calibrated from photometry, while relations calibrated from direct distance measurement are in better agreement., Comment: Accepted for publication in A&A

Classical Wolf-Rayet stars are descendants of massive OB-type stars that have lost their hydrogen-rich envelopes, and are in the final stages of stellar evolution, possibly exploding as type Ib/c supernovae. It is understood that the mechanisms driving this mass-loss are either strong stellar winds and or binary interactions, so intense studies of these binaries including their evolution can tell us about the importance of the two pathways in WR formation. WR 138 (HD 193077) has a period of just over 4 years and was previously reported to be resolved through interferometry. We report on new interferometric data combined with spectroscopic radial velocities in order to provide a three-dimensional orbit of the system. The precision on our parameters tend to be about an order of magnitude better than previous spectroscopic techniques. These measurements provide masses of the stars, namely $M_{\rm WR} = 13.93\pm1.49M_{\odot}$ and $M_{\rm O} = 26.28\pm1.71M_{\odot}$. The derived orbital parallax agrees with the parallax from \textit{Gaia}, namely with a distance of 2.13 kpc. We compare the system's orbit to models from BPASS, showing that the system likely may have been formed with little interaction but could have formed through some binary interactions either following or at the start of a red supergiant phase, but with the most likely scenario occurring as the red supergiant phase starts for a $\sim 40M_\odot$ star., Comment: accepted to ApJ

Classical Wolf-Rayet stars are the descendants of massive OB stars that have lost their hydrogen envelopes and are burning helium in their cores prior to exploding as type Ib/c supernovae. The mechanisms for losing their hydrogen envelopes are either through binary interactions or through strong stellar winds potentially coupled with episodic mass-loss. Amongst the bright classical WR stars, the binary system WR\,137 (HD\,192641; WC7d + O9e) is the subject of this paper. This binary is known to have a 13-year period and produces dust near periastron. Here we report on interferometry with the CHARA Array collected over a decade of time and providing the first visual orbit for the system. We combine these astrometric measurements with archival radial velocities to measure masses of the stars of $M_{\rm WR} = 9.5\pm3.4 M_\odot$ and $M_{\rm O} = 17.3\pm 1.9 M_\odot$ when we use the most recent \textit{Gaia} distance. These results are then compared to predicted dust distribution using these orbital elements, which match the observed imaging from \textit{JWST} as discussed recently by Lau et al. Furthermore, we compare the system to the BPASS models, finding that the WR star likely formed through stellar winds and not through binary interactions. However, the companion O star did likely accrete some material from the WR's mass-loss to provide the rotation seen today that drives its status as an Oe star., Comment: Accepted to ApJ

Processing large-scale graphs, containing billions of entities, is critical across fields like bioinformatics, high-performance computing, navigation and route planning, among others. Efficient graph partitioning, which divides a graph into sub-graphs while minimizing inter-block edges, is essential to graph processing, as it optimizes parallel computing and enhances data locality. Traditional in-memory partitioners, such as METIS and KaHIP, offer high-quality partitions but are often infeasible for enormous graphs due to their substantial memory overhead. Streaming partitioners reduce memory usage to O(n), where 'n' is the number of nodes of the graph, by loading nodes sequentially and assigning them to blocks on-the-fly. This paper introduces StreamCPI, a novel framework that further reduces the memory overhead of streaming partitioners through run-length compression of block assignments. Notably, StreamCPI enables the partitioning of trillion-edge graphs on edge devices. Additionally, within this framework, we propose a modification to the LA-vector bit vector for append support, which can be used for online run-length compression in other streaming applications. Empirical results show that StreamCPI reduces memory usage while maintaining or improving partition quality. For instance, using StreamCPI, the Fennel partitioner effectively partitions a graph with 17 billion nodes and 1.03 trillion edges on a Raspberry Pi, achieving significantly better solution quality than Hashing, the only other feasible algorithm on edge devices. StreamCPI thus advances graph processing by enabling high-quality partitioning on low-cost machines.

Social media has a significant impact on people's lives. Hate speech on social media has emerged as one of society's most serious issues in recent years. Text and pictures are two forms of multimodal data that are distributed within articles. Unimodal analysis has been the primary emphasis of earlier approaches. Additionally, when doing multimodal analysis, researchers neglect to preserve the distinctive qualities associated with each modality. To address these shortcomings, the present article suggests a scalable architecture for multimodal hate content detection called transformer-based multilevel attention (STMA). This architecture consists of three main parts: a combined attention-based deep learning mechanism, a vision attention-mechanism encoder, and a caption attention-mechanism encoder. To identify hate content, each component uses various attention processes and handles multimodal data in a unique way. Several studies employing multiple assessment criteria on three hate speech datasets such as Hateful memes, MultiOff, and MMHS150K, validate the suggested architecture's efficacy. The outcomes demonstrate that on all three datasets, the suggested strategy performs better than the baseline approaches.

Social media, particularly Twitter, has seen a significant increase in incidents like trolling and hate speech. Thus, identifying hate speech is the need of the hour. This paper introduces a computational framework to curb the hate content on the web. Specifically, this study presents an exhaustive study of pre-processing approaches by studying the impact of changing the sequence of text pre-processing operations for the identification of hate content. The best-performing pre-processing sequence, when implemented with popular classification approaches like Support Vector Machine, Random Forest, Decision Tree, Logistic Regression and K-Neighbor provides a considerable boost in performance. Additionally, the best pre-processing sequence is used in conjunction with different ensemble methods, such as bagging, boosting and stacking to improve the performance further. Three publicly available benchmark datasets (WZ-LS, DT, and FOUNTA), were used to evaluate the proposed approach for hate speech identification. The proposed approach achieves a maximum accuracy of 95.14% highlighting the effectiveness of the unique pre-processing approach along with an ensemble classifier.

The rapid advancement of machine learning technologies raises questions about the security of machine learning models, with respect to both training-time (poisoning) and test-time (evasion, impersonation, and inversion) attacks. Models performing image-related tasks, e.g. detection, and classification, are vulnerable to adversarial attacks that can degrade their performance and produce undesirable outcomes. This paper introduces a novel technique for anomaly detection in images called 2DSig-Detect, which uses a 2D-signature-embedded semi-supervised framework rooted in rough path theory. We demonstrate our method in adversarial settings for training-time and test-time attacks, and benchmark our framework against other state of the art methods. Using 2DSig-Detect for anomaly detection, we show both superior performance and a reduction in the computation time to detect the presence of adversarial perturbations in images.

We present the Capsule Vision 2024 Challenge: Multi-Class Abnormality Classification for Video Capsule Endoscopy. It was virtually organized by the Research Center for Medical Image Analysis and Artificial Intelligence (MIAAI), Department of Medicine, Danube Private University, Krems, Austria in collaboration with the 9th International Conference on Computer Vision & Image Processing (CVIP 2024) being organized by the Indian Institute of Information Technology, Design and Manufacturing (IIITDM) Kancheepuram, Chennai, India. This document provides an overview of the challenge, including the registration process, rules, submission format, description of the datasets used, qualified team rankings, all team descriptions, and the benchmarking results reported by the organizers., Comment: 11 pages

MIRC-X and MYSTIC are six-telescope near-infrared beam (1.08-2.38 ${\mu}$m) combiners at the CHARA Array on Mt Wilson CA, USA. Ever since the commissioning of MIRC-X (J and H bands) in 2018 and MYSTIC (K bands) in 2021, they have been the most popular and over-subscribed instruments at the array. Observers have been able to image stellar objects with sensitivity down to 8.1 mag in H and 7.8 mag in K-band under the very best conditions. In 2022 MYSTIC was upgraded with a new ABCD mode using the VLTI/GRAVITY 4-beam integrated optics chip, with the goal of improving the sensitivity and calibration. The ABCD mode has been used to observe more than 20 T Tauri stars; however, the data pipeline is still being developed. Alongside software upgrades, we detail planned upgrades to both instruments in this paper. The main upgrades are: 1) Adding a motorized filter wheel to MIRC-X along with new high spectral resolution modes 2) Updating MIRC-X optics to allow for simultaneous 6T J+H observations 3) Removing the warm window between the spectrograph and the warm optics in MYSTIC 4) Adding a 6T ABCD mode to MIRC-X in collaboration with CHARA/SPICA 5) Updating the MIRC-X CRED-ONE camera funded by Prof. Kraus from U. Exeter 6) Carrying out science verification of the MIRC-X polarization mode 7) Developing new software for ABCD-mode data reduction and more efficient calibration routines. We expect these upgrades to not only improve the observing experience, but also increase the sensitivity by 0.4 mag in J+H-bands, and 1 mag in K-band., Comment: 11 pages, 9 figures, Submitted to SPIE in Yokohama 2024

In this paper, we study an affine connection approach to realizing nonholonomic mechanical systems mediated by viscous friction forces with large coefficients, viewed as a singular perturbation of the nonholonomic system. We show that the associated slow manifold is represented coordinate-free as the image of a section over the nonholonomic distribution. We propose a novel invariance condition based on covariant derivatives and prove that this condition is equivalent to the classical invariance condition based on time derivatives. Accordingly, we propose a novel recursive procedure to approximate the slow manifold based on the covariant derivatives of a formal power series expansion of the section. Using this recurrence relation, we derive, up to second order, approximations of the slip velocities residing in the slow manifold, as well as the associated approximated dynamics up to first order. Lastly, we illustrate our approach with a case study of a vertical rolling disk., Comment: 30 pages

We continue our study of exponent semigroups of rational matrices. Our main result is that the matricial dimension of a numerical semigroup is at most its multiplicity (the least generator), greatly improving upon the previous upper bound (the conductor). For many numerical semigroups, including all symmetric numerical semigroups, our upper bound is tight.

Retrieval Augmented Generation (RAG) is widely used to enable Large Language Models (LLMs) perform Question Answering (QA) tasks in various domains. However, RAG based on open-source LLM for specialized domains has challenges of evaluating generated responses. A popular framework in the literature is the RAG Assessment (RAGAS), a publicly available library which uses LLMs for evaluation. One disadvantage of RAGAS is the lack of details of derivation of numerical value of the evaluation metrics. One of the outcomes of this work is a modified version of this package for few metrics (faithfulness, context relevance, answer relevance, answer correctness, answer similarity and factual correctness) through which we provide the intermediate outputs of the prompts by using any LLMs. Next, we analyse the expert evaluations of the output of the modified RAGAS package and observe the challenges of using it in the telecom domain. We also study the effect of the metrics under correct vs. wrong retrieval and observe that few of the metrics have higher values for correct retrieval. We also study for differences in metrics between base embeddings and those domain adapted via pre-training and fine-tuning. Finally, we comment on the suitability and challenges of using these metrics for in-the-wild telecom QA task., Comment: Accepted for publication in ICML 2024 Workshop on Foundation Models in the Wild

Cardiovascular disease is a major life-threatening condition that is commonly monitored using electrocardiogram (ECG) signals. However, these signals are often contaminated by various types of noise at different intensities, significantly interfering with downstream tasks. Therefore, denoising ECG signals and increasing the signal-to-noise ratio is crucial for cardiovascular monitoring. In this paper, we propose a deep learning method that combines a one-dimensional convolutional layer with transformer architecture for denoising ECG signals. The convolutional layer processes the ECG signal by various kernel/patch sizes and generates an embedding called multi-scale patch embedding. The embedding then is used as the input of a transformer network and enhances the capability of the transformer for denoising the ECG signal.

ESO's Very Large Telescope Interferometer has a history of record-breaking discoveries in astrophysics and significant advances in instrumentation. The next leap forward is its new visitor instrument, called Asgard. It comprises four natively collaborating instruments: HEIMDALLR, an instrument performing both fringe tracking and stellar interferometry simultaneously with the same optics, operating in the K band; Baldr, a Strehl optimizer in the H band; BIFROST, a spectroscopic combiner to study the formation processes and properties of stellar and planetary systems in the Y-J-H bands; and NOTT, a nulling interferometer dedicated to imaging nearby young planetary systems in the L band. The suite is in its integration phase in Europe and should be shipped to Paranal in 2025. In this article, we present details of the alignment and calibration unit, the observing modes, the integration plan, the software architecture, and the roadmap to completion of the project.

Our aim in this paper is to initiate the study of exponent semigroups for rational matrices. We prove that every numerical semigroup is the exponent semigroup of some rational matrix. We also obtain lower bounds on the size of such matrices and discuss the related class of power-integral matrices., Comment: 12 pages

A plethora of sentence embedding models makes it challenging to choose one, especially for technical domains rich with specialized vocabulary. In this work, we domain adapt embeddings using telecom, health and science datasets for question answering. We evaluate embeddings obtained from publicly available models and their domain-adapted variants, on both point retrieval accuracies, as well as their (95\%) confidence intervals. We establish a systematic method to obtain thresholds for similarity scores for different embeddings. As expected, we observe that fine-tuning improves mean bootstrapped accuracies. We also observe that it results in tighter confidence intervals, which further improve when pre-training is preceded by fine-tuning. We introduce metrics which measure the distributional overlaps of top-$K$, correct and random document similarities with the question. Further, we show that these metrics are correlated with retrieval accuracy and similarity thresholds. Recent literature shows conflicting effects of isotropy on retrieval accuracies. Our experiments establish that the isotropy of embeddings (as measured by two independent state-of-the-art isotropy metric definitions) is poorly correlated with retrieval performance. We show that embeddings for domain-specific sentences have little overlap with those for domain-agnostic ones, and fine-tuning moves them further apart. Based on our results, we provide recommendations for use of our methodology and metrics by researchers and practitioners.

Numerous recent X-ray observations of coronal loops in both active regions (ARs) and solar flares have shown clearly that elemental abundances vary with time. Over the course of a flare, they have been found to move from coronal values towards photospheric values near the flare peak, before slowly returning to coronal values during the gradual phase. Coronal loop models typically assume that the elemental abundances are fixed, however. In this work, we introduce a time-variable abundance factor into the 0D ebtel++ code that models the changes due to chromospheric evaporation in order to understand how this affects coronal loop cooling. We find that for strong heating events ($\gtrsim$ 1 erg s$^{-1}$ cm$^{-3}$), the abundances quickly tend towards photospheric values. For smaller heating rates, the abundances fall somewhere between coronal and photospheric values, causing the loop to cool more quickly than the time-fixed photospheric cases (typical flare simulations) and more slowly than time-fixed coronal cases (typical AR simulations). This suggests heating rates in quiescent AR loops no larger than $\approx 0.1$ erg s$^{-1}$ cm$^{-3}$ to be consistent with recent measurements of abundance factors $f \gtrsim 2$., Comment: Accepted to ApJL. Comments and criticisms are more than welcome! The paper is fully reproducible, with scripts to run the simulations and generate the figures available on this Github repository: https://github.com/jwreep/ebtel_abundances

Large Language Models (LLMs) have achieved state-of-the-art performance at zero-shot generation of abstractive summaries for given articles. However, little is known about the robustness of such a process of zero-shot summarization. To bridge this gap, we propose relevance paraphrasing, a simple strategy that can be used to measure the robustness of LLMs as summarizers. The relevance paraphrasing approach identifies the most relevant sentences that contribute to generating an ideal summary, and then paraphrases these inputs to obtain a minimally perturbed dataset. Then, by evaluating model performance for summarization on both the original and perturbed datasets, we can assess the LLM's one aspect of robustness. We conduct extensive experiments with relevance paraphrasing on 4 diverse datasets, as well as 4 LLMs of different sizes (GPT-3.5-Turbo, Llama-2-13B, Mistral-7B, and Dolly-v2-7B). Our results indicate that LLMs are not consistent summarizers for the minimally perturbed articles, necessitating further improvements., Comment: Accepted to NAACL 2025 Findings