A reasoning model's emitted text trace offers at best a sanitized, post-hoc rationalization of its actual internal computation. To peek behind the curtain, researchers Kang Chen, Sihan Zhao, Yixin Cao, and Yugang Jiang from Fudan University and the Shanghai Innovation Institute have introduced a two-level internal readout framework that decodes latent reasoning dynamics directly from mixture-of-experts (MoE) architectures.
Rather than parsing tokens, the team distilled vocabulary-scale J-space into J64—a compressed 64-axis semantic frame extracted straight from hidden states. J64 isolates latent cognitive shifts that never surface in the visible output, cleanly distinguishing between routine forward-pass verbosity and genuine problem-induced computational strain. On benchmark evaluations, the framework delivers a solid 0.096 to 0.135 boost in held-out AUC compared to baseline methods that merely track token occupancy.
Crucially, this visibility does not demand crushing compute overhead. The researchers reconstructed the semantic frame directly from native expert routing statistics into a lightweight proxy, dubbed R64. Across three models and two architecture families, R64 maintained a median per-axis correlation of 0.69 to 0.86 with J64 while preserving 95% to 100% of its predictive lift on gpt-oss-20b.
In practice, tapping raw routing telemetry allows engineers to detect whether an inference path is actively converging or simply spinning in a degenerate reasoning loop long before the final tokens drop. Dynamic test-time selection policies driven by these rolling readouts boosted accuracy by up to 5.9 points, effectively turning low-level MoE routing signals into a real-time, hardware-level control plane for complex inference workloads.