LatentHarness Unifies Memory and Reasoning via Latent Action Selection

LatentHarness treats memory access and reasoning as a sequential latent action selection problem using THINK, RECALL, and EXIT actions. It utilizes counterfactual policy distillation to train models to decide when to retrieve evidence versus when to perform further computation.