Target speaker extraction using microphone arrays is a crit- ical task in speech processing in complex acoustic environ- ments. In this paper, we propose an informed neural filter- ing approach that incorporates a priori source information into multi-level feature representations. The network adopts a simple yet effective insertion strategy in which each block maintains uniform dimensionality and semantic consistency, enabling the a priori information to be integrated through a unified mechanism. Within this framework, three represen- tative speaker-embedding fusion strategies are investigated; notably, two of them adaptively refine the a priori source in- formation using shallow features derived from array obser- vations. Experimental results demonstrate the effectiveness of both multi-level information injection and adaptive refine- ment of a priori information.