RSRoy Schwartz

Roy Schwartz

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Roy Schwartz has a total of 207 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Do OCTA morphological patterns in neovascular AMD actually matter? A scoping review of clinical utility and terminological overlap

    by Daniela Bacherini, Ken Kawamoto, Gianni Virgili, Clara Rizzo, Caroline R. Baumal, Usha Chakravarthy, Christine A. Curcio, Livia Faes, K. Bailey Freund, David Huang, et al.

    Purpose: Optical coherence tomography angiography (OCTA) has enabled detailed in vivo imaging of macular neovascularization (MNV) in neovascular age-related macular degeneration (nAMD), leading to a proliferation of morphological descriptive terms. This scoping review aimed to systematically map the existing literature on OCTA-based MNV morphology and evaluate its correlation with disease activity. Methods: A systematic literature search covering publications through 2025 was performed. After screening 2,445 titles and obtaining 60 full texts, 43 studies were included. Data on morphological terminology, study design, and correlation with disease activity were extracted and synthesized. Results: The included studies, encompassing a total of 2,712 eyes, identified a vast and heterogeneous lexicon of qualitative descriptors, including terms such as “medusa,” “sea-fan,” and “glomerulus”, characterized by significant terminological overlap and inconsistent definitions across studies. Inconsistent, low-certainty associations between specific OCTA morphologies and MNV activity were identified across studies, with findings limited by significant methodological heterogeneity. The evidence for these patterns as reliable biomarkers of disease activity is weak and often contradictory, and several studies were found to use OCTA features to define activity, creating circular arguments that undermine their conclusions. Conclusion: The current terminology for OCTA morphology in nAMD is fragmented and inconsistently applied. The link between these patterns and disease activity is poorly established, severely limiting their clinical utility. These findings highlight a need for a standardized, consensus-based framework for describing and interpreting OCTA findings in nAMD.

  • Open Access

    Reticular pseudodrusen load is associated with an increased risk of stroke

    by Roy Schwartz, Abraham Olvera-Barrios, Alasdair N. Warwick, Hagar Khalid, Sumita Phatak, Mahima Jhingan, Coen de Vente, Philippe Valmaggia, Sandra Liakopoulos, Yanling Ouyang, et al.

    Background/aims: Reticular pseudodrusen (RPD) are increasingly recognised as a distinct phenotype in the age-related macular degeneration (AMD) disease spectrum. This study investigates the association between RPD and cardiovascular outcomes, specifically myocardial infarction (MI) and stroke in the UK Biobank (UKBB). Methods: Retrospective analysis of UKBB participants (n=2010). A validated deep learning framework identified and quantified subjects with RPD, drusen and controls on optical coherence tomography. Five retina specialists validated the artificial intelligence findings. Multivariable logistic regression models assessed the association between RPD/drusen and stroke, MI and combined MI/stroke, adjusting for age, sex, high-density lipoprotein cholesterol/low-density lipoprotein cholesterol ratio and smoking history. Results: The study cohort included 71 subjects with pure RPD, 401 with pure drusen, 368 with both and 1170 controls. Univariable analysis showed that for every 20 RPD lesions, the OR for stroke increased by 1.02 (95% CI 1.00 to 1.04, p=0.028). The mean number of RPD per patient in this cohort was 199, indicating an increased stroke risk of 20%. Multivariable analysis showed a significant association between RPD load and stroke risk (OR 1.02, 95% CI 1.00 to 1.04, p=0.042) after adjustment for confounders. No significant association was found between drusen and stroke or between RPD/drusen and MI. Conclusion: Increased RPD load is associated with a higher risk of stroke, independent of traditional cardiovascular risk factors. Drusen did not have similar associations, suggesting that RPD may represent a distinct disease entity. Evaluating cardiovascular risk in patients with numerous RPD might be advisable. Further prospective studies are needed to validate these findings and explore the underlying mechanisms.

  • Open Access

    SpeLLM: Character-Level Multi-Head Decoding

    by Amit Ben-Artzy, Roy Schwartz

    Scaling LLM vocabulary is often used to reduce input sequence length and alleviate attention's quadratic cost. Yet, current LLM architectures impose a critical bottleneck to this procedure: the output projection layer scales linearly with vocabulary size, rendering substantial expansion impractical. We propose SpeLLM, a method that decouples input and output vocabularies by predicting character-level strings through multiple output heads. In SpeLLM, each of the $k$ linear heads predicts a single character simultaneously, enabling the model to represent a much larger output space using smaller, independent linear heads. We present a self-distillation approach for converting a standard LLM to a SpeLLM. Our experiments with four pre-trained LLMs show their SpeLLM variants achieve competitive performance on downstream tasks while reducing runtime by 5.1% on average across models. Our approach provides a potential avenue for reducing LLM costs, while increasing support for underrepresented languages and domains.

  • Open Access

    Genetic Distinctions Between Reticular Pseudodrusen and Drusen: A Genome-Wide Association Study

    by Roy Schwartz, Alasdair N. Warwick, Anthony P. Khawaja, Robert Luben, Hagar Khalid, Sumita Phatak, Mahima Jhingan, Coen de Vente, Philippe Valmaggia, Sandra Liakopoulos, et al.


  • Open Access

    On Pruning State-Space LLMs

    by Tamer Ghattas, Michael Hassid, Roy Schwartz

    Tamer Ghattas, Michael Hassid, Roy Schwartz. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.


  • Open Access

    Genetic Distinctions Between Reticular Pseudodrusen and Drusen: Insights from a Genome-Wide Association Study

    by Roy Schwartz, Alasdair N. Warwick, Anthony P. Khawaja, Robert Luben, Hagar Khalid, Sumita Phatak, Mahima Jhingan, Coen de Vente, Philippe Valmaggia, Sandra Liakopoulos, et al.

    Purpose: To explore genetic determinants specific to reticular pseudodrusen (RPD), and to compare these with genetic associations for drusen. Setting: Participants with RPD, drusen, or controls from the UK Biobank (UKBB), a large, multisite, community-based cohort study. Methods: A previously validated deep learning framework was deployed on 169,370 optical coherence tomography (OCT) volumes from the UKBB to identify cases with RPD and/or drusen and controls without these phenotypes. Cases were included if they were 60 years or older, with at least 5 lesions. Five retina specialists manually validated the cohorts using OCT and color fundus photographs. Quantifications of RPD and drusen derived by the framework were used as variables. Two primary genome-wide association study (GWAS) analyses were performed to explore potential genetic associations with number of RPD and drusen within ‘pure’ cases, where only RPD or drusen were present in either eye. A candidate approach was furthermore adopted to assess 46 previously known AMD loci. Secondary GWAS were undertaken for number of RPD and drusen in mixed cases, as well as binary case-control analyses for pure RPD and pure drusen. Genome-wide significance was defined as pARMS2-HTRA1 (p=1.9e-09), rs79641866/ PARD3B on chromosome 2 (p=1.3e-08), rs143184903/ ITPR1 on chromosome 3 (p=8.1e-09), and rs76377757/ SLN on chromosome 11 (p=4.3e-08). The latter three are all uncommon variants (minor allele frequency CFH was also observed adopting a candidate approach (p=1.8e-04). Two loci reached genome-wide significance for the primary pure drusen GWAS: rs10801555/ CFH on chromosome 1 (p=6.0e-33) and rs61871744/ ARMS2-HTRA1 on chromosome 10 (p=4.2e-20). For the mixed RPD and drusen secondary analyses, lead variants at both the CFH and ARMS2-HTRA1 loci reached genome-wide significance, with C2-CFB-SKIV2L additionally associated for mixed drusen alone. Findings from the binary case-control GWAS for drusen mirrored those of the primary drusen analysis, however no variants reached genome-wide significance in the case-control RPD GWAS. Conclusions: Our findings indicate a clear association between the ARMS2 - HTRA1 locus with higher RPD load. Although the association at the CFH locus did not reach genome-wide significance, we observed a suggestive link. We furthermore identified three novel associations that are unique to RPD, albeit for uncommon genetic variants. These findings were only observable when quantifying RPD load as a continuous trait, which increased the study’s statistical power. Further studies with larger sample sizes are now required to explore the relative contributions of AMRS2-HTRA1 and CFH to RPD development, and to explore the validity of these newly presented RPD-specific genetic associations.


  • Open Access

    Beyond Performance: Quantifying and Mitigating Label Bias in LLMs

    by Yuval Reif, Roy Schwartz

    Large language models (LLMs) have shown remarkable adaptability to diverse tasks, by leveraging context prompts containing instructions, or minimal input-output examples. However, recent work revealed they also exhibit label bias -- an undesirable preference toward predicting certain answers over others. Still, detecting and measuring this bias reliably and at scale has remained relatively unexplored. In this study, we evaluate different approaches to quantifying label bias in a model's predictions, conducting a comprehensive investigation across 279 classification tasks and ten LLMs. Our investigation reveals substantial label bias in models both before and after debiasing attempts, as well as highlights the importance of outcomes-based evaluation metrics, which were not previously used in this regard. We further propose a novel label bias calibration method tailored for few-shot prompting, which outperforms recent calibration approaches for both improving performance and mitigating label bias. Our results emphasize that label bias in the predictions of LLMs remains a barrier to their reliability.


  • Open Access

    Attend First, Consolidate Later: On the Importance of Attention in Different LLM Layers

    by Amit Ben Artzy, Roy Schwartz

    Amit Ben Artzy, Roy Schwartz. Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP. 2024.


  • Open Access

    Beyond Performance: Quantifying and Mitigating Label Bias in LLMs

    by Yuval Reif, Roy Schwartz

    Yuval Reif, Roy Schwartz. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.


  • Open Access

    Breaking Common Sense: WHOOPS! A Vision-and-Language Benchmark of Synthetic and Compositional Images

    by Nitzan Bitton-Guetta, Yonatan Bitton, Jack Hessel, Ludwig Schmidt, Yuval Elovici, Gabriel Stanovsky, Roy Schwartz

    Weird, unusual, and uncanny images pique the curiosity of observers because they challenge commonsense. For example, an image released during the 2022 world cup depicts the famous soccer stars Lionel Messi and Cristiano Ronaldo playing chess, which playfully violates our expectation that their competition should occur on the football field. Humans can easily recognize and interpret these unconventional images, but can AI models do the same? We introduce WHOOPS!, a new dataset and benchmark for visual commonsense. The dataset is comprised of purposefully commonsense-defying images created by designers using publicly-available image generation tools like Midjourney. We consider several tasks posed over the dataset. In addition to image captioning, cross-modal matching, and visual question answering, we introduce a difficult explanation generation task, where models must identify and explain why a given image is unusual. Our results show that state-of-the-art models such as GPT3 and BLIP2 still lag behind human performance on WHOOPS!. We hope our dataset will inspire the development of AI models with stronger visual commonsense reasoning abilities. Data, models and code are available at the project website: whoops-benchmark.github.io