arXiv subjects

Aditi Gupta

Publications and source records attributed to Aditi Gupta.

13 recordsLinked to original sources

arXiv2026

Test-Time Training for Modality Order Consistency in Vision-Language Models

We find that vision-language models are sensitive to a specific semantically irrelevant change: the order in which the image and question are presented. Across three models and three benchmarks, image first prompting consistently outperforms question-first prompting, revealing a repeatable modality order failure. We use this gap to design an order-consistent test-time training method. Our method substantially closes the modality-order gap across all evaluated settings. Surprisingly, it also yields consistent improvements in the stronger image-first branch over the baseline, hence bootstrapping both orderings toward mutual consistency. Activation patching localizes the ordering failure to a narrow mid-network region where representations diverge sharply between prompt orders. We find that the test-time training method repairs this misalignment across layers. Together, our results identify modality-order sensitivity as a circuit-level failure in VLMs and demonstrate that simple, asymmetric test-time adaptation can effectively mitigate it and even improve performance over the baseline.

cs.CV↗

arXiv2026

Speculative Refinement: A Hybrid Autoregressive Diffusion Decoding Strategy and Its Behavior Across Benchmarks

How should we evaluate generation systems that combine autoregressive (AR) and diffusion decoding? We study this question through Speculative Refinement (SpecRef), a training-free hybrid method that warm-starts a masked diffusion language model from an AR draft using entropy-guided selective masking. Evaluating SpecRef across six benchmarks (HumanEval, MBPP, GSM8K, BBH, ARC-Challenge, HellaSwag) with three distinct evaluation protocols (execution-based pass@1, exact-match, log-likelihood scoring), we surface several findings relevant beyond our specific system: (1) code benchmarks conflate structural discovery with logical correctness: providing a syntactic scaffold lifts accuracy from near zero to over 20% without changing the model, indicating that much of the baseline failure is structural; (2) a refinement tension phenomenon where multi-stage correction degrades already-correct tokens, exposing benchmark saturation ceilings invisible to single-model evaluation; (3) log-likelihood and generative evaluation produce different model rankings for the same model pair, suggesting they measure different capabilities; (4) standard Python post-processing silently breaks code evaluation for non-AR generators. These observations apply to any multi-stage or non-autoregressive generation pipeline and point toward more diagnostic evaluation practices.

cs.SE↗

arXiv2026

Sharpen Your Flow: Sharpness-Aware Sampling for Flow Matching

Flow matching models generate samples by numerically integrating a learned velocity field, with each integration step requiring a neural network evaluation. Fast generation therefore requires using a small fixed evaluation budget effectively: the key question is not only how to integrate the flow, but where the sampler should spend its steps. We propose SharpEuler, a training-free sampler that profiles a pretrained model offline by estimating where the learned velocity field changes most rapidly along calibration trajectories. This finite-difference estimate defines a solver-aware sharpness profile, which is smoothed and converted by a quantile transform into a timestep grid for any desired inference budget. At test time, sampling remains ordinary Euler integration with the same number of model evaluations as a uniform schedule. We justify SharpEuler using three principles: a numerical principle identifying trajectory acceleration as the leading source of Euler discretization error, a variational principle deriving sharpness-based power-law timestep densities, and a statistical guarantee showing that the finite-sample calibrated sampler is stable at the terminal distribution level. Our experiments show that SharpEuler improves sample quality at fixed budgets, reducing inter-mode leakage and increasing mode coverage.

cs.LG↗

arXiv2026

Quantifying Epistemic Uncertainty in Diffusion Models

To ensure high quality outputs, it is important to quantify the epistemic uncertainty of diffusion models. Existing methods are often unreliable because they mix epistemic and aleatoric uncertainty. We introduce a method based on Fisher information that explicitly isolates epistemic variance, producing more reliable plausibility scores for generated data. To make this approach scalable, we propose FLARE (Fisher-Laplace Randomized Estimator), which approximates the Fisher information using a uniformly random subset of model parameters. Empirically, FLARE improves uncertainty estimation in synthetic time-series generation tasks, achieving more accurate and reliable filtering than other methods. Theoretically, we bound the convergence rate of our randomized approximation and provide analytic and empirical evidence that last-layer Laplace approximations are insufficient for this task.

stat.ML↗

arXiv2025

Hierarchical Sparse Plus Low Rank Compression of LLM

Modern large language models (LLMs) place extraordinary pressure on memory and compute budgets, making principled compression indispensable for both deployment and continued training. We present Hierarchical Sparse Plus Low-Rank (HSS) compression, a two-stage scheme that (i) removes the largest-magnitude weights into a sparse matrix S and (ii) applies a recursive Hierarchically Sparse Separable (HSS) low-rank factorisation to the dense residual matrix. A recursive rank-reducing strategy and a reverse Cuthill-Mckee (RCM) permutation are introduced to align high weights towards the diagonal with the block-diagonal hierarchy, maximising off-diagonal compressibility (because they are touched only once). HSS is hardware-friendly: its matrix-vector multiply reduces to one sparse and a sequence of thin-matrix multiplications and can be trained end-to-end with standard optimisers. Experiments on LLaMA-7B show that targeting only the self-attention projections (1.6 B parameters of Q, K, and V matrices out of a total 7B parameters) suffices to yield large memory savings while retaining comparable state-of-the-art perplexity scores on test samples of the WikiText dataset. For example, with a 30\% sparsity budget and an outer rank of 512, sHSS-RCM achieves a perplexity of 1.64, outperforming dense baselines and classical sparse-plus-SVD variants, while also achieving significant memory savings.

cs.LG↗

arXiv2025

Global PIQA: Evaluating Commonsense Reasoning Across 100+ Languages and Cultures

To date, there exist almost no culturally-specific evaluation benchmarks for large language models (LLMs) that cover a large number of languages and cultures. In this paper, we present Global PIQA, a participatory commonsense reasoning benchmark for over 100 languages, constructed by hand by over 350 researchers from over 65 countries around the world. The 141 language varieties in Global PIQA cover five continents, 19 language families, and 24 writing systems. In the non-parallel split of Global PIQA, over 50% of examples reference local foods, customs, traditions, or other culturally-specific elements. In the parallel split, we translate more "culturally agnostic" commonsense reasoning questions into 131 language varieties, for direct cross-lingual comparisons. In both splits, all examples have been verified by native speakers of the languages. We find that state-of-the-art LLMs perform well on Global PIQA in aggregate, but they exhibit weaker performance in lower-resource languages (e.g. up to a 68% accuracy gap between languages in the parallel split). Global PIQA highlights that in many languages and cultures, everyday knowledge remains an area for improvement in LLMs, alongside more widely-discussed capabilities such as complex reasoning and expert knowledge. Beyond its uses for LLM evaluation, Global PIQA provides a glimpse into the wide diversity of cultures in which human language is embedded.

cs.CL↗

arXiv2024

Optical Coherence Tomography Angiography-OCTA dataset for the study of Diabetic Retinopathy

This study presents a dataset consisting of 268 retinal images from 179 individuals, including 133 left-eye and 135 right-eye images, collected from Natasha Eye Care and Research Institute in Pune, Maharashtra, India. The images were captured using a nonmydriatic Optical Coherence Tomography Angiography (OCTA) device, specifically the Optovue Avanti Edition machine as per the protocol mentioned in this paper. Two ophthalmologists then annotated the images. This dataset can be used by researchers and doctors to develop automated diagnostic tools for early detection of diabetic retinopathy (DR).

eess.IV↗

arXiv2024

Lumbar Spine Tumor Segmentation and Localization in T2 MRI Images Using AI

In medical imaging, segmentation and localization of spinal tumors in three-dimensional (3D) space pose significant computational challenges, primarily stemming from limited data availability. In response, this study introduces a novel data augmentation technique, aimed at automating spine tumor segmentation and localization through AI approaches. Leveraging a fusion of fuzzy c-means clustering and Random Forest algorithms, the proposed method achieves successful spine tumor segmentation based on predefined masks initially delineated by domain experts in medical imaging. Subsequently, a Convolutional Neural Network (CNN) architecture is employed for tumor classification. Moreover, 3D vertebral segmentation and labeling techniques are used to help pinpoint the exact location of the tumors in the lumbar spine. Results indicate a remarkable performance, with 99% accuracy for tumor segmentation, 98% accuracy for tumor classification, and 99% accuracy for tumor localization achieved with the proposed approach. These metrics surpass the efficacy of existing state-of-the-art techniques, as evidenced by superior Dice Score, Class Accuracy, and Intersection over Union (IOU) on class accuracy metrics. This innovative methodology holds promise for enhancing the diagnostic capabilities in detecting and characterizing spinal tumors, thereby facilitating more effective clinical decision-making.

eess.IV↗

arXiv2021

Machine learning for modeling the progression of Alzheimer disease dementia using clinical data: a systematic literature review

Objective Alzheimer disease (AD) is the most common cause of dementia, a syndrome characterized by cognitive impairment severe enough to interfere with activities of daily life. We aimed to conduct a systematic literature review (SLR) of studies that applied machine learning (ML) methods to clinical data derived from electronic health records in order to model risk for progression of AD dementia. Materials and Methods: We searched for articles published between January 1, 2010, and May 31, 2020, in PubMed, Scopus, ScienceDirect, IEEE Explore Digital Library, Association for Computing Machinery Digital Library, and arXiv. We used predefined criteria to select relevant articles and summarized them according to key components of ML analysis such as data characteristics, computational algorithms, and research focus. Results: There has been a considerable rise over the past 5 years in the number of research papers using ML-based analysis for AD dementia modeling. We reviewed 64 relevant articles in our SLR. The results suggest that majority of existing research has focused on predicting progression of AD dementia using publicly available datasets containing both neuroimaging and clinical data (neurobehavioral status exam scores, patient demographics, neuroimaging data, and laboratory test values). Discussion: Identifying individuals at risk for progression of AD dementia could potentially help to personalize disease management to plan future care. Clinical data consisting of both structured data tables and clinical notes can be effectively used in ML-based approaches to model risk for AD dementia progression. Data sharing and reproducibility of results can enhance the impact, adaptation, and generalizability of this research.

q-bio.QM↗

arXiv2016

Strong Selection Significantly Increases Epistatic Interactions in the Long-Term Evolution of a Protein

Epistatic interactions between residues determine a protein's adaptability and shape its evolutionary trajectory. When a protein experiences a changed environment, it is under strong selection to find a peak in the new fitness landscape. It has been shown that strong selection increases epistatic interactions as well as the ruggedness of the fitness landscape, but little is known about how the epistatic interactions change under selection in the long-term evolution of a protein. Here we analyze the evolution of epistasis in the protease of the human immunodeficiency virus type 1 (HIV-1) using protease sequences collected for almost a decade from both treated and untreated patients, to understand how epistasis changes and how those changes impact the long-term evolvability of a protein. We use an information-theoretic proxy for epistasis that quantifies the co-variation between sites, and show that positive information is a necessary (but not sufficient) condition that detects epistasis in most cases. We analyze the "fossils" of the evolutionary trajectories of the protein contained in the sequence data, and show that epistasis continues to enrich under strong selection, but not for proteins whose environment is unchanged. The increase in epistasis compensates for the information loss due to sequence variability brought about by treatment, and facilitates adaptation in the increasingly rugged fitness landscape of treatment. While epistasis is thought to enhance evolvability via valley-crossing early-on in adaptation, it can hinder adaptation later when the landscape has turned rugged. However, we find no evidence that the HIV-1 protease has reached its potential for evolution after 9 years of adapting to a drug environment that itself is constantly changing.

q-bio.PE↗

arXiv2015

Evolution of genome size in asexual populations

Genome sizes have evolved to vary widely, from 250 bases in viroids to 670 billion bases in amoeba. This remarkable variation in genome size is the outcome of complex interactions between various evolutionary factors such as point mutation rate, population size, insertions and deletions, and genome editing mechanisms that may be specific to certain taxonomic lineages. While comparative genomics analyses have uncovered some of the relationships between these diverse evolutionary factors, we still do not understand what drives genome size evolution. Specifically, it is not clear how primordial mutational processes of base substitutions, insertions, and deletions influence genome size evolution in asexual organisms. Here, we use digital evolution to investigate genome size evolution by tracking genome edits and their fitness effects in real time. In agreement with empirical data, we find that mutation rate is inversely correlated with genome size in asexual populations. We show that at low point mutation rate, insertions are significantly more beneficial than deletions, driving genome expansion and acquisition of phenotypic complexity. Conversely, high mutational load experienced at high mutation rates inhibits genome growth, forcing the genomes to compress genetic information. Our analyses suggest that the inverse relationship between mutation rate and genome size is a result of the tradeoff between evolving phenotypic innovation and limiting the mutational load.

q-bio.PE↗

arXiv2015

TweetCred: Real-Time Credibility Assessment of Content on Twitter

During sudden onset crisis events, the presence of spam, rumors and fake content on Twitter reduces the value of information contained on its messages (or "tweets"). A possible solution to this problem is to use machine learning to automatically evaluate the credibility of a tweet, i.e. whether a person would deem the tweet believable or trustworthy. This has been often framed and studied as a supervised classification problem in an off-line (post-hoc) setting. In this paper, we present a semi-supervised ranking model for scoring tweets according to their credibility. This model is used in TweetCred, a real-time system that assigns a credibility score to tweets in a user's timeline. TweetCred, available as a browser plug-in, was installed and used by 1,127 Twitter users within a span of three months. During this period, the credibility score for about 5.4 million tweets was computed, allowing us to evaluate TweetCred in terms of response time, effectiveness and usability. To the best of our knowledge, this is the first research work to develop a real-time system for credibility on Twitter, and to evaluate it on a user base of this size.

cs.CR↗

arXiv2012

Characterizing Pedophile Conversations on the Internet using Online Grooming

Cyber-crime targeting children such as online pedophile activity are a major and a growing concern to society. A deep understanding of predatory chat conversations on the Internet has implications in designing effective solutions to automatically identify malicious conversations from regular conversations. We believe that a deeper understanding of the pedophile conversation can result in more sophisticated and robust surveillance systems than majority of the current systems relying only on shallow processing such as simple word-counting or key-word spotting. In this paper, we study pedophile conversations from the perspective of online grooming theory and perform a series of linguistic-based empirical analysis on several pedophile chat conversations to gain useful insights and patterns. We manually annotated 75 pedophile chat conversations with six stages of online grooming and test several hypothesis on it. The results of our experiments reveal that relationship forming is the most dominant online grooming stage in contrast to the sexual stage. We use a widely used word-counting program (LIWC) to create psycho-linguistic profiles for each of the six online grooming stages to discover interesting textual patterns useful to improve our understanding of the online pedophile phenomenon. Furthermore, we present empirical results that throw light on various aspects of a pedophile conversation such as probability of state transitions from one stage to another, distribution of a pedophile chat conversation across various online grooming stages and correlations between pre-defined word categories and online grooming stages.

cs.CY↗