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

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.

Introduction

Chain-of-Thought (CoT) prompting has significantly enhanced the ability of Large Language Models (LLMs) to solve complex problems. However, the faithfulness of these reasoning chains is often questioned: is the model truly thinking through the steps, or is it merely providing a post-hoc justification for an answer it has already internally decided upon? This phenomenon, termed "Reasoning Theater," suggests that models may generate long, performative explanations that do not reflect their internal state, leading to wasted compute and reduced interpretability.

Methodology: Three Lenses of Belief

To investigate when a model actually commits to an answer, researchers employed three distinct methods to monitor internal and external progress:

  1. Attention Probes (White-box): Lightweight classifiers trained on the model’s internal residual streams. These probes pool information across the reasoning prefix to detect if the final answer is already encoded in hidden states.
  2. Forced Answering (Black-box): Truncating the reasoning at various points and forcing the model to provide an immediate multiple-choice answer.
  3. CoT Monitoring (External): Using an external LLM to read the generated text and predict the final answer based solely on the visible reasoning.
Overview of the three lenses used to evaluate model commitment: Attention Probes, Forced Answering, and CoT Monitoring.

Results: Performative vs. Genuine Reasoning

The study evaluated DeepSeek-R1 (671B) and GPT-OSS (120B) on two benchmarks: MMLU-Redux (general knowledge) and GPQA-Diamond (graduate-level science). The results revealed a stark contrast based on task difficulty.

On easier tasks (MMLU), models often exhibit "Reasoning Theater." Internal probes show high confidence in the correct answer almost from the first token, while the external CoT monitor remains at chance performance for much of the generation. This indicates the model "knows" the answer but continues to generate performative tokens.

Conversely, on difficult tasks (GPQA-Diamond), the theater disappears. The internal probes and external monitor gain confidence at similar rates, suggesting that the model genuinely requires the step-by-step computation to reach a conclusion.

Decoding accuracy across MMLU and GPQA-Diamond for DeepSeek-R1 and GPT-OSS, showing the gap between internal probes and external monitors.

Key Insights

  • Inflection Points are Honest: Verbalized shifts in logic, such as "Wait, let me rethink," correlate strongly with actual shifts in internal probe confidence, suggesting these moments are not faked.
  • The Incentive for Verbosity: Current Reinforcement Learning (RL) training may accidentally incentivize models to be verbose rather than pragmatic, encouraging long reasoning chains even when unnecessary.
  • Efficiency through Early Exit: By using attention probes as an "early exit" signal—stopping generation once internal confidence reaches a threshold—researchers achieved massive token savings. This method saved 80% of tokens on MMLU and 30% on GPQA-Diamond while maintaining approximately 97% of the original accuracy.

Conclusion

Reasoning models frequently engage in performative CoT for tasks they can solve via internal weights. By disentangling internal beliefs from external text, we can better understand model faithfulness and implement adaptive computation strategies. Probing internal activations offers a path toward more efficient and interpretable AI, allowing us to bypass "Reasoning Theater" when the model has already reached its conclusion.

Summary updated

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.
BibTeX
@inproceedings{boppana2026reasoning,
  author = {Siddharth Boppana and Annabel Ma and Max Loeffler and Raphaël Sarfati and Eric Bigelow and Atticus Geiger and Owen Lewis and Jack Merullo},
  title = {{Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thought}},
  booktitle = {International Conference on Machine Learning},
  year = {2026},
  url = {https://openreview.net/forum?id=1lyagkzogH},
}
Figure