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

Siddharth Boppana, Annabel Ma, Max Loeffler, Raphaël Sarfati, Eric Bigelow, Atticus Geiger, Owen Lewis, Jack Merullo
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:9152-9177, 2026.

Abstract

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.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-boppana26a, title = {Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thought}, author = {Boppana, Siddharth and Ma, Annabel and Loeffler, Max and Sarfati, Rapha\"{e}l and Bigelow, Eric and Geiger, Atticus and Lewis, Owen and Merullo, Jack}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {9152--9177}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/boppana26a/boppana26a.pdf}, url = {https://proceedings.mlr.press/v306/boppana26a.html}, abstract = {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.} }
Endnote
%0 Conference Paper %T Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thought %A Siddharth Boppana %A Annabel Ma %A Max Loeffler %A Raphaël Sarfati %A Eric Bigelow %A Atticus Geiger %A Owen Lewis %A Jack Merullo %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-boppana26a %I PMLR %P 9152--9177 %U https://proceedings.mlr.press/v306/boppana26a.html %V 306 %X 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.
APA
Boppana, S., Ma, A., Loeffler, M., Sarfati, R., Bigelow, E., Geiger, A., Lewis, O. & Merullo, J.. (2026). Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thought. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:9152-9177 Available from https://proceedings.mlr.press/v306/boppana26a.html.

Related Material