Figure 1: Overview of the BrainIPAgent framework. (1) The Multi-Agent Interpretation Frame-
work extracts structural patterns via collaborative reasoning to generate sparsity masks and textual
priors. (2) The Structure-Aware Functional Encoder maps these patterns to attention biases in
the logit space and projects them via mask propagation. (3) The RL-Guided Alignment Module
dynamically regulates structural guidance strength Alpha (α) and alignment weight Lambda (λ).