Decoding Hidden Reasoning States in Mixture-of-Experts Models with J64 and R64

An arXiv paper (2608.17638) presents a two-tier internal readout designed for mixture-of-experts (MoE) reasoning models. Initially, the approach condenses the vocabulary-scale J-space into J64, a 64-axis semantic framework derived from the model's reasoning states. This process uncovers state information that is not visible in the emitted trace, distinguishing inference effort from problem-induced strain. J64 achieves an increase of 0.096 to 0.135 in held-out AUC compared to a baseline that reads the same rollout as token occupancy. Subsequently, the team constructs R64 from the original expert-routing statistics, yielding a low-overhead proxy with median per-axis correlation ranging from 0.69 to 0.86 across two families and three models. On gpt-oss-20b, R64 maintains 95–100% of J64's predictive advantage, facilitating test-time decisions at two different temporal resolutions.