<p>Signage systems play a vital role in facilitating safe and efficient evacuation during emergencies in complex buildings. However, existing studies often rely on idealized assumptions and fail to capture individuals’ actual cognitive and perceptual responses to signage, resulting in discrepancies between simulation predictions and real-world evacuation behavior. To address this gap, this study constructs an experimental platform integrating virtual reality (VR) and eye-tracking technologies to objectively record participants’ visual attention distribution and signage recognition during evacuation scenarios. Experimental results indicate that the effectiveness of signage is significantly constrained by individual visual-cognitive mechanisms. Based on these findings, an improved multi-signage evacuation guidance model is developed, which comprehensively accounts for factors such as static and dynamic signage, fire-induced environmental changes, and guidance field intensity. The model further optimizes the effective influence range of signage. Simulation results show strong consistency with empirical data (<i>p</i> = 0.5244 &gt; 0.05), confirming the model’s validity and reliability. In addition, simulations under varying signage densities reveal that higher signage density does not always enhance evacuation efficiency; instead, there exists an optimal threshold, with the best guidance performance observed at a density of ρ = 2. This study provides theoretical foundations and empirical evidence for the scientific deployment of evacuation signage and the optimization of evacuation strategies, offering valuable insights for practical applications in building safety and emergency management.</p>