LLMs & Generative AI

← Back to Artificial Intelligence

Large language models, text generation, multimodal models, and prompt engineering.

This section explores the structural mechanics, scaling limits, and real-world applications of large language models and generative architectures. Our research investigates the entire lifecycle of these systems, beginning with the transition from pre-training and supervised fine-tuning to post-training alignment paradigms like direct preference optimization. We dissect the fundamental constraints of transformer architectures, examining how subword tokenization algorithms introduce cognitive blind spots—such as the "strawberry problem"—and how long-context systems optimize memory bandwidth through techniques like speculative decoding, latent attention, and key-value cache management.

Beyond core training, this research analyzes the evolution of prompt engineering from basic linguistic prompting to structured systems architecture, including agentic workflows, chain-of-thought processing, and advanced reasoning topologies. These methodologies are weighed against the inherent limitations of next-token prediction, detailing why AI models exhibit hallucinations, fail at analogical reasoning, and suffer from biases incentivized by flawed evaluation benchmarks. This structural analysis extends into multimodality, mapping how latent diffusion models, generative adversarial networks, and early-fusion architectures enable image, video, music, and de novo protein generation.

In practical application, we evaluate how these models perform in high-stakes environments. This includes the implementation of LLMs in quantitative finance—covering alpha generation, trade execution latency, and secure fine-tuning on proprietary data using parameter-efficient methods and differential privacy—as well as their utility in legal reasoning. Finally, our research addresses the cognitive and societal footprint of generative technology. We examine the tension between human learning and automated assistance, exploring how cognitive offloading and epistemic debt alter human writing skills, educational assessment frameworks, and long-term creative capacity.

40 published articles