Evolutionary multi-task optimization enhances overall search efficiency by sharing implicit genetic information among distinct optimization tasks. However, when inter-task correlation is low, blind knowledge transfer without correlation measurement is highly prone to triggering negative transfer effects. To address this issue, this paper proposes a multifactorial evolutionary algorithm based on dynamic gradient similarity (MFEA-DGS). Firstly, a quasi-gradient operator based on directional gradient descent is introduced to construct descent direction vectors, enabling the algorithm to capture the geometric features of the solution space for efficient positive transfer. Secondly, a dynamic gradient similarity strategy is proposed to adaptively regulate the intensity of knowledge transfer and physically block conflicting tasks using a hard-threshold mechanism, by calculating the directional cosine similarity between the gradient vectors of the source and target tasks. Additionally, a probabilistic hybrid search framework combining the gradient operator with simulated binary crossover and polynomial mutation is constructed. This framework leverages the global exploration capability of evolutionary operators to compensate for the limitations of pure gradient search, which is prone to falling into local optima, thereby significantly improving convergence precision. Experimental results on the CEC2017-MTSO and WCCI2020-MTSO benchmark suites demonstrate that the MFEA-DGS achieves superior performance in terms of convergence speed, solution accuracy, and robustness compared to several state-of-the-art multitasking algorithms.