<p>Argument mining is a fundamental task in natural language understanding that aims to identify the structure of argumentative discourse units within a corpus. However, existing approaches often struggle with rigid dependencies on pre-segmented spans (oracle boundaries) and lack the flexibility to adapt to diverse argumentative ontologies. The primary objective of this study is to establish a robust, end-to-end pipeline that autonomously detects argument boundaries, types, and relations without relying on unrealistic priors. To this end, we propose Argus, a novel framework that aims to integrate autoregressive Large Language Models (LLMs) via a unique mark-based generation strategy. Unlike conventional classification or structured-output methods, Argus uses a schema-agnostic prompting template that enables it to generalize across heterogeneous polemic structures and ontologies while preserving the natural distribution of the text. We efficiently implement the training pipeline using Parameter-Efficient Fine-Tuning (PEFT) and Low-Rank Adaptation (LoRA) to handle the computational constraints of LLMs. Empirical evaluations on benchmark datasets demonstrate that our model outperforms competitive baselines, particularly in realistic scenarios where argument boundaries are not predefined.</p>