prompt

Free

Route reasoning workloads by evidence: early-exit theater, deep-reason substance.

FreeFree tier

Inputs: textOutputs: text

Type

Open Source

Visit Website GitHub (40 stars)

About prompt

The Reasoning Theater Diagnostician is a system prompt designed to determine whether a language model's chain-of-thought (CoT) reasoning is genuine or performative 'theater.' Based on research from 'Reasoning Theater: Disentangling Model Beliefs from CoT' (arXiv 2603.05488, 2026), it distinguishes tasks where the final answer is already encoded in early-layer activations before any CoT tokens are emitted (theater) from tasks where CoT actively shifts the model's beliefs (substance). The prompt prescribes a routing policy that applies probe-guided early-exit on simple tasks to reduce token generation by up to 80% with no accuracy loss, while ensuring deep reasoning on hard tasks where CoT is necessary.

Key Features

Classifies chain-of-thought as theater or substance per (model, task, prompt template) triple

Probe-guided early-exit reduces token generation by up to 80% on simple tasks with no accuracy cost

Empirically grounded, repeatable, and reversible audit per workload

Designs routing policy that allocates CoT budget only to workloads that need it

Explicit uncertainty bands at the boundary between theater and substance

Avoids early-exit on genuine-CoT tasks where belief shifts occur during reasoning

Pros & Cons

Pros

Can reduce token generation by up to 80% on simple tasks at no accuracy cost
Empirically grounded; not philosophical or speculative
Per-triple audit is repeatable, reversible, and measurable
Clear design philosophy with non-negotiable empirical basis

Cons

Requires access to early-layer activations, which may not be available in all model architectures or APIs
Boundary cases with uncertain bands may require additional overhead to resolve
Effectiveness depends on the specific model and prompt template; may need calibration per triple

Best For

Optimizing inference costs by early-exiting on simple reasoning tasksImproving LLM efficiency without sacrificing accuracyAuditing model behavior to understand genuine vs. performative reasoningResearch on chain-of-thought mechanisms and model beliefsBuilding adaptive inference systems that scale compute based on task difficulty

FAQ

What is reasoning theater?

Reasoning theater occurs when a model's final answer is already decodable from its early-layer activations before the chain-of-thought has emitted a single token. The subsequent CoT tokens are performative—they do not change the model's belief.

When should early-exit be applied?

Early-exit should be applied only on tasks where the answer is pre-decided (theater). It must never be applied on tasks where CoT produces genuine belief shifts (substance), as that would sacrifice accuracy.

How is the theater/substance classification determined?

The classification is per (model, task, prompt template) triple and is repeatable and reversible. It involves auditing early-layer activations to see if the final answer can be decoded before CoT begins, and comparing that to the final answer after full CoT.