College of San Mateo

Large language models are widely used but expensive to run. To reduce costs, it is crucial to maximize request batch size through efficient GPU memory management. Existing approaches, such as PagedAttention, struggle with modern LLMs because of the growing heterogeneity in the sizes of models' internal embeddings and attention mechanisms.

Aamvata is a chronic, systemic inflammatory disorder rooted in the vitiation of Vata dosha alongside the accumulation of Ama (undigested toxic matter). It corresponds to conditions like Rheumatoid Arthritis in modern medicine. Rasonadi Kwath, a classical decoction mentioned in Bhaishajya Ratnavali, is highly esteemed in Ayurvedic practice for the treatment of Aamvata due to its potent Deepana-Pachana (digestive), Vatahara, and Shothahara (anti-inflammatory) properties. This formulation integrates potent herbs such as Rasona (Garlic), Pippali, Haritaki, and others, delivering a comprehensive therapeutic effect. The present review elaborates on the ingredients, preparation method, pharmacological attributes, and mechanism of action based on classical texts and modern interpretations, to establish its relevance in Aamvata Chikitsa and encourage further scientific validation.

Reinforcement Learning with Human Feedback (RLHF) has emerged as a key paradigm for task-specific fine-tuning of language models using human preference data.While numerous publicly available preference datasets provide pairwise comparisons of responses, the potential for biases in the resulting reward models remains underexplored.In this work, we introduce novel methods to detect and evaluate prefix bias-a systematic shift in model preferences triggered by minor variations in query prefixes-in LLM-based reward models trained on such datasets.We leverage these metrics to reveal significant biases in preference models across racial and gender dimensions.Our comprehensive evaluation spans diverse open-source preference datasets and reward model architectures, demonstrating susceptibility to this kind of bias regardless of the underlying model architecture.Furthermore, we propose a data augmentation strategy to mitigate these biases, showing its effectiveness in reducing the impact of prefix bias.Our findings highlight the critical need for bias-aware dataset design and evaluation in developing fair and reliable reward models, contributing to the broader discourse on fairness in AI.

Abstract Introduction: Despite digitization of clinical data, drug approval processes remain prohibitively slow and 80% of trials miss enrollment timelines. Clinical enrollment and completion are influenced by both protocol design and operational factors - yet quantitative evidence linking these factors to trial outcomes remains scarce. This study developed a novel framework leveraging large language models (LLMs) and machine learning to quantify key factors such as protocol and trial complexity to assess impact on trial completion. Using a dataset of 12,096 interventional oncology trials from ClinicalTrials.gov, we transformed unstructured trial protocols into machine-readable formats, extracting key metrics such as eligibility criteria, cohort size, enrollment windows, and adherence to clinical guidelines. Methods: Eligibility criteria were abstracted and standardized using advanced machine learning (ML) techniques, including large language models (LLMs), entity recognition (NER), and unsupervised clustering using the unstructured text of trial protocols as well as guidelines from the Food and Drug Administration (FDA) and National Clinical Trials Network (NCTN). This resulted in a comprehensive database of protocol features such as eligibility criteria, interventions, endpoints, and operational details. These features were further categorized into key clinical and study domains such as cancer diagnosis, prior therapy, comorbidities, and concomitant medications. A Random Forest classifier, trained on 60% of the dataset with rigorous cross-validation, predicted trial completion with an accuracy of 81.36% (±1.39%), precision of 79.10% (±1.63%), and F1 score of 80.87% (±1.96%). Model performance was consistent across trial phases, with Phase 2 trials demonstrating the highest accuracy (82.42% ± 2.57%). Sensitivity analyses with gradient boosting and deep learning validated these findings, though the Random Forest model's interpretability was preferred. Results: Key predictors of trial success included larger target cohort size, longer recruitment windows, and number of recruitment sites, all positively correlated with completion rates. Conversely, higher protocol complexity—measured by the total number of eligibility criteria, inclusion/exclusion counts, and deviations from clinical research guidelines—was negatively associated with successful trial completion. Simplified protocols with fewer restrictive criteria and alignment with recommended guidelines significantly improved outcomes. Conclusion: This study provides a scalable, flexible framework for optimizing clinical trial design. By identifying and quantifying critical protocol features, this approach enhances trial efficiency, promotes alignment with best practices, and offers actionable insights to improve enrollment rates and patient outcomes. The methodology is extensible, enabling prospective applications to refine future trial protocols across diverse therapeutic areas. Citation Format: Ognjen Nikolic, Pranav Singh, Anne-Marie Meyer, Shannon Silkensen, Carla Rodriguez-Watson, Richard L. Schilsky. Structuring the unstructured: How protocol quality and precision drive clinical trial completion [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 2 (Late-Breaking, Clinical Trial, and Invited Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_2):Abstract nr LB373.

Although an increased risk of the skin cancer melanoma in people with Parkinson's disease (PD) has been shown in multiple studies, the mechanisms involved are poorly understood, but increased expression of the PD-associated protein alpha-synuclein (αSyn) in melanoma cells may be important. Our previous work suggests that αSyn can facilitate DNA double-strand break (DSB) repair, promoting genomic stability. We now show that αSyn is preferentially enriched within the nucleolus in melanoma, where it colocalizes with DNA damage markers and DSBs. Inducing DSBs specifically within nucleolar ribosomal DNA (rDNA) increases αSyn levels near sites of damage. αSyn knockout increases DNA damage within the nucleolus at baseline, after specific rDNA DSB induction, and prolongs the rate of recovery from this induced damage. αSyn is important downstream of ataxia-telangiectasia-mutated signaling to facilitate MDC1-mediated 53BP1 recruitment to DSBs, reducing micronuclei formation and promoting cellular proliferation, migration, and invasion.

In light of the accelerated growth of e-commerce, the generation of high-quality and efficient content has emerged as a pivotal factor in enhancing user experience and business value. However, conventional methods of content creation are prone to inefficiencies and creative constraints. This paper puts forth a comprehensive model based on the combination of existing generative pre-trained models and human-computer interaction design, with the objective of enhancing the efficiency and quality of e-commerce content generation. In particular, the initial stage of the process involves the utilisation of multi-modal data fusion technology, which facilitates the integration of diverse input sources, including images, videos, and textual data. This approach enhances the model's capacity to comprehend the multifaceted nature of products and services, thereby optimising the generation process. Secondly, a cross-domain sentiment analysis and recommendation engine was designed, which combined the self-attention mechanism in deep learning with the objective of analysing the popularity of the generated content according to consumer behaviour in real time and adjusting the generation strategy. Furthermore, the real-time feedback loop mechanism has been innovatively introduced, which enables the dynamic optimisation of content generation in accordance with user evaluation and click-through rate. This facilitates an improvement in the personalisation and accuracy of content. The experimental results demonstrate that the model markedly enhances the efficiency of content generation and user satisfaction on the e-commerce platform.