From Neural Network Decisions to Training Cases: An Exact Account via Case-Based Decision Theory

This paper shows that an OLS action readout fitted on a fixed neural representation admits an exact case-based decomposition: each action score becomes a weighted sum of training-case returns, with coefficients determined by empirical Gram geometry. The authors distinguish a regime where these weights support CBDT-style similarity semantics from the general case, where they should be interpreted as signed geometric influence. Evaluations on synthetic CBDT, PJM, Adult Income, and Default Credit tasks report strong Top-30 consistency and competitive support reconstruction. The audit requires only an OLS probe and score-reconstruction checks, without retraining the representation or recovering the original optimization trajectory.