Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thought

We provide evidence of \textit{performative} chain-of-thought (CoT) in reasoning models, where a model becomes strongly confident in its final answer, but continues generating tokens without revealing its internal belief. Our analysis compares activation probing, early forced answering, and a CoT monitor across two large models (DeepSeek-R1 671B & GPT-OSS 120B) and find task difficulty-specific differences: The model's final answer is decodable from activations far earlier in CoT than a monitor is able to say, especially for easy recall-based MMLU questions. We contrast this with genuine reasoning in difficult multihop GPQA-Diamond questions. Despite this, \textit{inflection points} (e.g., backtracking, `aha' moments) occur almost exclusively in responses where probes show large belief shifts, suggesting these behaviors track genuine uncertainty rather than learned ``reasoning theater." Finally, probe-guided early exit reduces tokens by up to 80\% on MMLU and 30\% on GPQA-Diamond with similar accuracy, positioning attention probing as an efficient tool for detecting performative reasoning and enabling adaptive computation.

Reasoning models often “think out loud” using chain-of-thought (CoT), but it is unclear whether these explanations always reflect the model’s true internal reasoning. In this work, we show evidence of performative CoT: situations where a model has already internally settled on an answer with high confidence, yet continues generating reasoning tokens without revealing that certainty. We study this behavior in two large reasoning models, DeepSeek-R1 671B and GPT-OSS 120B, using three complementary tools: activation probing, early forced answering, and a CoT monitor. Across benchmarks ranging from easier recall-style MMLU questions to challenging multihop GPQA-Diamond problems, we find that the model’s final answer can often be decoded from its internals much earlier than its generated reasoning text suggests. This gap is especially large on simpler questions, where models appear to continue “reasoning” after already knowing the answer. At the same time, we find evidence that not all CoT behavior is merely performance. Distinctive reasoning events such as backtracking and sudden “aha” moments are strongly associated with large shifts in the model’s internal beliefs, indicating that these behaviors often reflect genuine uncertainty and computation on harder tasks. Finally, we show that internal probes can be used to stop generation early once the model’s answer becomes stable. This probe-guided early exit cuts token usage by up to 80% on MMLU and 30% on GPQA-Diamond while maintaining similar accuracy. Our results suggest that activation probing can both help identify performative reasoning and enable more efficient adaptive inference in large reasoning models.