Reasoning Theater: Performative Chain-of-Thought in Reasoning Models

arXiv 2603.05488·high signal

Boppana, Ma, Loeffler et al. demonstrate that reasoning models (DeepSeek-R1 671B, GPT-OSS 120B) engage in "performative" CoT — the model's final answer is decodable from activations far earlier than the visible chain-of-thought suggests, especially on easy tasks (MMLU). Genuine belief shifts (backtracking, "aha" moments) only occur on hard problems (GPQA-Diamond). Activation probing enables adaptive computation: up to 80% token reduction on MMLU and 30% on GPQA-Diamond with similar accuracy. Cri