Financial entity-level sentiment analysis aims to identify sentiment polarity toward specific financial entities in financial texts. This task is challenging because a single sentence may contain multiple entities with opposite sentiment orientations, while financial sentiment is often conveyed through domain-specific and implicit cues, increasing the risk of entity confusion. To address these challenges, we propose MSAR, a Multi-Stage Adversarial Representation Learning framework built on a FinBERT-CRF backbone. At the textual level, a Model-Sensitive Metric generates adversarial examples by jointly considering model sensitivity, semantic consistency, and linguistic fluency. At the representation level, Adaptive Adversarial Regularization dynamically adjusts the regularization strength according to gradient sensitivity and output-distribution variation. By integrating discrete textual perturbation with continuous representation regularization, MSAR enhances robust entity-level sentiment prediction. Experiments on FinEntity and SentFin show that MSAR outperforms representative baselines. Ablation, sensitivity, and low-resource experiments further confirm the effectiveness and robustness of the proposed components. These findings demonstrate that jointly optimizing textual and representation-level robustness improves financial entity-level sentiment analysis, particularly under low-resource conditions.