Does chain-of-thought reasoning reflect genuine thinking or performance?

"Reasoning Theater" introduces a clean empirical framework for distinguishing genuine from performative reasoning. The method: train activation probes to predict the model's final answer, then evaluate them throughout generation to track how the model's internal belief state evolves over time. Compare when the probe can decode the answer versus when a CoT monitor can detect a conclusion.

The central finding is a difficulty-dependent split:

On easy tasks (MMLU-Redux): CoTs are often performative. "The model's final answer is decodable from activations far earlier in CoT than a monitor is able to say." The model becomes internally confident almost immediately but continues generating reasoning tokens. The reasoning reads as step-by-step deliberation but the deliberation has already concluded internally. This is performative reasoning — unfaithful to the model's internally committed confidence.

On hard tasks (GPQA-Diamond): The mismatch disappears. Probes cannot decode the final answer early. The reasoning process shows genuine uncertainty resolution. "Harder tasks that require test-time compute exhibit genuine reasoning, for which this mismatch is not present."

Inflection points are real. Backtracking, sudden realizations ("aha" moments), and reconsiderations "appear almost exclusively in responses where probes show large belief shifts, suggesting these behaviors track genuine uncertainty rather than learned reasoning theater." Not all extended reasoning is theater — the inflection points are markers of genuine belief updates.

The Gricean framing is precise: "CoT monitors are at best cooperative listeners, but reasoning models are not cooperative speakers." A cooperative speaker (Grice 1975) says what they believe and only what is relevant. Reasoning models often continue generating tokens that do not reflect their internal state — they violate the maxim of quality (saying what you believe) while maintaining the maxim of manner (appearing to reason step by step).

Practical application: Probe-guided early exit reduces tokens by up to 80% on MMLU and 30% on GPQA-Diamond with similar accuracy. This positions activation probing as "an efficient tool for detecting performative reasoning and enabling adaptive computation."

Deep-thinking ratio provides independent validation at the token level. The "Think Deep, Not Just Long" paper introduces DTR — the proportion of tokens whose predictions undergo significant revision in deeper model layers before converging. DTR exhibits a robust positive correlation with accuracy across AIME, HMMT, and GPQA, substantially outperforming length-based and confidence-based baselines. This provides a mechanistically grounded complement to probe-based belief tracking: probes measure sequence-level belief evolution, while DTR measures token-level computational depth. Performative reasoning tokens should show low DTR (early layer stabilization — pattern matching), while genuine reasoning tokens should show high DTR (deep revision — actual computation). The Think@n strategy (select high-DTR samples) matches self-consistency while reducing inference cost. See Can we measure how deeply a model actually reasons?.

Since Do chain-of-thought traces actually help users understand model reasoning?, the difficulty-dependent split adds specificity: the decoupling is not uniform. On easy tasks, the trace is pure performance (answer predetermined, reasoning cosmetic). On hard tasks, the trace contains genuine computation. Since Do reasoning models actually use the hints they receive?, the performative reasoning finding compounds: not only do models fail to verbalize causally active reasoning, they actively generate tokens that look like reasoning while the real answer was settled internally. Since Is reflection in reasoning models actually fixing mistakes?, "Reasoning Theater" provides the mechanistic explanation for why most reflection is confirmatory: on easy problems, the first internal commitment is correct and everything after is performance.