arXiv subjects

Sanjay Adhikesaven

Publications and source records attributed to Sanjay Adhikesaven.

4 recordsLinked to original sources

arXiv2026

Which Models Are Our Models Built On? Auditing Invisible Dependencies in Modern LLMs

Modern LLM training pipelines increasingly rely on other models to generate data, filter corpora, judge outputs, and guide development decisions. These dependencies are recursive: a model may depend on an upstream artifact whose own dependencies are documented only in separate releases and artifacts. As a result, the full dependency structure is fragmented across heterogeneous public artifacts, with complexity and recursive depth far outpacing humans' ability to trace. We introduce ModSleuth, an agentic system that recursively reconstructs LLM dependency graphs from public artifacts with source-grounded evidence. We find that the primary challenge is no longer information extraction, but defining what constitutes a dependency and reconciling artifact references across inconsistent documentation. We address these challenges through a formalization that distinguishes direct and indirect dependencies, represents heterogeneous pipeline roles through operation-centered relationships, and resolves artifact identities across names, versions, and repositories. Applying ModSleuth to four public-artifact-rich LLM releases, we recover 1,060 source-verified dependencies and construct large-scale dependency graphs of modern LLM development. These graphs reveal multi-hop license obligations, train-evaluation coupling, discrepancies between released and training-time artifacts, and documentation inconsistencies that would otherwise be difficult to uncover. We release ModSleuth and the resulting dependency graphs to support transparent analysis of the increasingly complex ecosystems underlying modern LLMs.

cs.CL↗

arXiv2026

Train Separately, Merge Together: Modular Post-Training with Mixture-of-Experts

Extending a fully post-trained language model with new domain capabilities is fundamentally limited by monolithic training paradigms: retraining from scratch is expensive and scales poorly, while continued training often degrades existing capabilities. We present BAR (Branch-Adapt-Route), which trains independent domain experts, each through its own mid-training, supervised finetuning, and reinforcement learning pipeline, and composes them via a Mixture-of-Experts architecture with lightweight router training. Unlike retraining approaches that mix all domains and require full reprocessing for any update (with cost scaling quadratically), BAR enables updating individual experts independently with linear cost scaling and no degradation to existing domains. At the 7B scale, with experts for math, code, tool use, and safety, BAR achieves an overall score of 49.1 (averaged across 7 evaluation categories), matching or exceeding re-training baselines (47.8 without mid-training, 50.5 with). We further show that modular training provides a structural advantage: by isolating each domain, it avoids the catastrophic forgetting that occurs when late-stage RL degrades capabilities from earlier training stages, while significantly reducing the cost and complexity of updating or adding a domain. Together, these results suggest that decoupled, expert-based training is a scalable alternative to monolithic retraining for extending language models.

cs.LG↗

arXiv2022

An Industrial Workplace Alerting and Monitoring Platform to Prevent Workplace Injury and Accidents

Workplace accidents are a critical problem that causes many deaths, injuries, and financial losses. Climate change has a severe impact on industrial workers, partially caused by global warming. To reduce such casualties, it is important to proactively find unsafe environments where injuries could occur by detecting the use of personal protective equipment (PPE) and identifying unsafe activities. Thus, we propose an industrial workplace alerting and monitoring platform to detect PPE use and classify unsafe activity in group settings involving multiple humans and objects over a long period of time. Our proposed method is the first to analyze prolonged actions involving multiple people or objects. It benefits from combining pose estimation with PPE detection in one platform. Additionally, we propose the first open source annotated data set with video data from industrial workplaces annotated with the action classifications and detected PPE. The proposed system can be implemented within the surveillance cameras already present in industrial settings, making it a practical and effective solution.

cs.CV↗

arXiv2022

Challenges of Equitable Vaccine Distribution in the COVID-19 Pandemic

The COVID-19 pandemic has led to a need for widespread and rapid vaccine development. As several vaccines have recently been approved for human use or are in different stages of development, governments across the world are preparing comprehensive guidelines for vaccine distribution and monitoring. In this early article, we identify challenges in logistics, health outcomes, user-centric matters, and communication associated with disease-related, individual, societal, economic, and privacy consequences. Primary challenges include difficulty in equitable distribution, vaccine efficacy, duration of immunity, multi-dose adherence, and privacy-focused record-keeping to be HIPAA compliant. While many of these challenges have been previously identified and addressed, some have not been acknowledged from a comprehensive view accounting for unprecedented interactions between challenges and specific populations. The logistics of equitable widespread vaccine distribution in disparate populations and countries of various economic, racial, and cultural constitutions must be thoroughly examined and accounted for. We also describe unique challenges regarding the efficacy of vaccines in specialized populations including children, the elderly, and immunocompromised individuals. Furthermore, we report the potential for understudied drug-vaccine interactions as well as the possibility that certain vaccine platforms may increase susceptibility to HIV. Given these complicated issues, the importance of privacy-focused, user-centric systems for vaccine education and incentivization along with clear communication from governments, organizations, and academic institutions is imperative. These challenges are by no means insurmountable, but require careful attention to avoid consequences spanning a range of disease-related, individual, societal, economic, and security domains.

econ.GN↗