On the Bias of Next-Token Predictors Toward Systematically Inefficient Reasoning: A Shortest-Path Case StudyPublished in NeurIPS 25, 2025Share on Bluesky Facebook LinkedIn X (formerly Twitter) Previous Next