Reproduce “Beyond the Trace: Coupling an Interpretable Reasoning-State Readout to Native MoE Routing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kang Chen, Sihan Zhao, Yixin Cao, Yu-Gang Jiang
The authors introduce J64, which compresses vocabulary-aligned readings of mixture-of-experts hidden states into 64 named coordinates, and R64, which reconstructs them from routing statistics without a second forward pass. On gpt-oss-20b at High effort, adding J64 raises held-out trajectory-outcome AUC from 0.730 to 0.827. The evidence centers on competition math and does not establish general controllability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.