Researchers have developed a new method to read out the internal reasoning states of mixture-of-experts (MoE) models. This allows for the extraction of readable process state from the model's own reasoning states, which is not shown in its emitted trace. The method, called J64, reveals how inference effort and problem-induced strain are separated. It also improves held-out AUC by 0.096 to 0.135 over a baseline. Additionally, a low-overhead proxy called R64 can reconstruct J64 from native expert-routing statistics, with a median per-axis correlation of 0.69 to 0.86 across three models and two families. --- Why it matters: This matters because it allows researchers to understand the internal workings of MoE models, which are increasingly being used in applications such as natural language processing. By making latent process state readable, J64 can help improve model accuracy and decision-making at test time. Source: https://arxiv.org/abs/2608.17638