Bio-Inspired Adaptive Dynamic Attention: An Empirically Driven AI Framework for Human–Machine Coaching in Team Collaborative Decision-Making

Abstract

This paper introduces the Dynamic Cognitive Load-LSTM Attention Routing (DCLAR) framework for real-time AI-driven coaching. We combine real-time indicators of mental workload with an attention-based model to coordinate team collaboration. Moving beyond static multi-head attention models, we present a Gated Cognitive Load Estimator (GCLE) that leverages physiological and behavioral signals—including heart rate variability and speech rate—to infer participants cognitive load in real time.We use the load values to decide which attention heads should stay active in the LSTM at each step, enhancing computational efficiency without compromising critical contextual information. A residual gating mechanism is further incorporated to fuse attention outputs with LSTM hidden states, ensuring stable gradient propagation amid cognitive load fluctuations. Implemented on edge devices such as the NVIDIA Jetson Orin, DCLAR operates with sub-millisecond latency. Experimental evaluations demonstrate a reduction of up to 40% in redundant computations compared to static benchmarks, while maintaining comparable performance. By linking findings from cognitive science to model design, we create an AI coach that adapts to users’ mental states.