Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thought
This work addresses the problem of inefficient and potentially misleading CoT generation for researchers and practitioners working with large language models, by providing a method to detect when models are performing 'reasoning theater' versus genuine reasoning.
This paper investigates 'performative chain-of-thought' (CoT) in large language models, where models generate CoT despite already having a confident final answer. They found that the final answer is decodable from activations much earlier than indicated by the CoT, especially for easy MMLU questions, and that probe-guided early exiting can reduce tokens by up to 80% on MMLU and 30% on GPQA-Diamond with similar accuracy.
We provide evidence of 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, 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.