Abstract: The integration of mobile edge computing (MEC) with intelligent connected vehicles (ICVs) presents an innovative framework to address resource-constrained environments prevalent in the Internet of Vehicles (IoV). As ICV technology continues to evolve, emerging compute-intensive applications—such as autonomous driving navigation and augmented reality interfaces—are driving the demand for enhanced real-time processing capabilities and resource allocation. However, the inherent limitations of onboard computing resources in ICVs pose significant challenges for the effective execution of these latency-sensitive applications. Although existing research on computation offloading has made progress, two major limitations persist: 1) insufficient consideration of intertask dependencies in complex applications and 2) overly simplistic assumptions that primarily focus on single-application scenarios, thereby overlooking heterogeneous multi-application environments typically utilized by ICVs. To address these gaps, this paper proposes a deep reinforcement learning (DRL)-based partial offloading algorithm specifically designed for MEC-driven ICV scenarios, where tasks exhibit directed acyclic graph (DAG) dependencies across multiple applications. The proposed method employs a two-stage hierarchical modeling architecture. In the first stage, by leveraging dependency-aware scheduling, dynamic execution priorities are assigned to convert the complex DAG topology into a linear task chain. In the second stage, a heterogeneous DAG workflow aggregation strategy is introduced, transforming the multi-DAG offloading problem into a unified single-DAG optimization framework to enable efficient resource coordination across concurrent applications. To model the offloading decision process, the system is formalized as a Markov decision process (MDP), where each state transition corresponds to a binary offloading decision (local execution versus edge server offloading), effectively balancing latency and energy consumption. The solution to the MDP is formulated using a sequence-to-sequence neural network architecture with hierarchical recurrent layers. This architecture captures the spatiotemporal dependencies between subtasks by encoding historical task states with a bidirectional gated recurrent unit and utilizing an attention mechanism in the decoder to predict the optimal offloading operation. Furthermore, the system adopts the Asynchronous Advantage Actor-Critic (A3C) algorithm, which integrates parallel exploration to enhance policy diversity and improve training efficiency. By deploying multiple agents with shared neural parameters, the A3C framework ensures faster convergence by reducing the variance in gradient updates while maintaining a comprehensive exploration of the state-action space. The experimental results validate the effectiveness of the proposed algorithm. Compared with baseline methods in single-DAG, multi-DAG, and real IoV edge environments, the overall utility of the proposed method demonstrates a significant improvement. Specifically, the method achieves an enhancement of 3.2–8.7% in the latency energy tradeoff in real-world scenarios by leveraging the asynchronous update mechanism of parallel agents. In addition, as the density of edge servers increases, the algorithm dynamically adjusts to balance the computational load, outperforming complete offloading and random strategies by 25.1–34.7%. These results confirm that the proposed algorithm effectively coordinates asynchronous computing and dynamic communication constraints, providing a reliable solution for balancing latency and energy consumption in heterogeneous ICV applications. In conclusion, the DRL-based partial offloading algorithm proposed in this study effectively addresses the challenges associated with task dependencies and multi-application environments in MEC-driven ICVs. It demonstrates significant advantages in optimizing resource allocation and enhancing the overall system performance, positioning it as a promising solution for next-generation intelligent transportation systems.