AI Failure Modes Shift to Incoherence as Reasoning Complexity Increases

New research from Anthropic Fellows decomposes frontier model errors into systematic bias and random variance, finding that longer reasoning chains lead to failures dominated by incoherence, or 'hot mess' behavior, rather than pure goal misalignment.