Reproduce “Emergent Manifold Separability during Reasoning in Large Language Models” 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
Chanwoo Chun, Alexandre Polo, SueYeon Chung
The paper uses Manifold Capacity Theory to track how linearly separable node answers are in a model’s residual stream during chain-of-thought reasoning. Across Boolean and prose-based eligibility tasks, capacity rises when an answer is computed or reused, then falls. Probes can retain above-90% accuracy as capacity nears baseline. On depth-5 Boolean tasks, CoT reached 99.2% final-answer accuracy, within the reported tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.