On the Bias of Next-Token Predictors Toward Systematically Inefficient Reasoning: AShortest-Path Case Study

Open in new window