Attention Trajectories as a Diagnostic Axis for Deep Reinforcement Learning
The paper "Attention Trajectories as a Diagnostic Axis for Deep Reinforcement Learning" analyzes how AI agents learn by tracking their internal attention patterns over time, rather than just looking at final performance. Through case studies involving Pong and biomechanical models, it demonstrates that these patterns reveal inherent biases. This method allows researchers to diagnose vulnerabilities and design more reliable, transparent AI systems.