Lihao Sun @1e0sun.bsky.social · 14/04/2026Each reasoning step occupies a distinct region in representation space, and these regions become increasingly linearly separable with layer depth. This structure generalizes across tasks and answer formats. 2/ 110 0
Lihao Sun @1e0sun.bsky.social · 14/04/2026Trajectories for correct and incorrect reasoning start out nearly identical but diverge systematically at late steps. Late-step trajectory features predict final-answer correctness with AUC 0.87 - before the answer is generated. 3/ 110 0
Lihao Sun @1e0sun.bsky.social · 14/04/2026This step-specific organization exists across training regimes and is already present in base models. From this perspective, reasoning training doesn't introduce new representational organization - it accelerates convergence toward a termination-related subspace. 4/ 110 0
