Interpretable Reasoning Readout for MoE Models Improves Performance

This research introduces a novel approach to gain deeper insights into the internal reasoning processes of Mixture-of-Experts (MoE) models, going beyond just the emitted output. The method involves a two-level internal readout system. First, it distills the model's complex reasoning states into a compact, interpretable 64-axis semantic frame, termed J64. This J64 readout can reveal crucial process information, such as inference effort versus problem-induced strain, which is not evident from the model's output trace alone. The second level involves reconstructing this J64 frame directly from the native expert-routing statistics within the MoE architecture, resulting in a low-overhead proxy called R64. This R64 proxy maintains a high correlation with J64 and preserves most of its predictive gains, making it practical for deployment. The readout system supports real-time decision-making during inference. Both J64 and R64 have been shown to improve single-branch selection and voting mechanisms over candidate sets. During text generation, a rolling readout window enables a stop-and-resample policy, significantly boosting accuracy. Furthermore, targeted edits to the router, guided by J64, can induce desired reasoning behaviors, shifting models from numerical guessing towards more precise symbolic execution.