Research map / Large language models / Reasoning and chain of thought

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

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Studies performative chain-of-thought in DeepSeek-R1 671B and GPT-OSS 120B by comparing attention probes, forced answering, and a Gemini CoT monitor, and uses probe-guided early exit to save tokens.

Problem

Whether reasoning models' final answers are already internally decodable well before their chain-of-thought reveals them (performative CoT), and whether this differs by task difficulty.

Approach

Trains attention probes over residual-stream activations at every prefix position to decode the eventual answer, and compares their early decoding against forced answering and a CoT monitor, using the gap as a measure of performativity; probe confidence also serves as an early-exit signal.

Key results

• Probe-guided early exit at 95% confidence on MMLU retains 97% of original performance while saving 68% of tokens; at 80% confidence on GPQA-Diamond saves 50% of tokens with 97% performance retained.
• Performativity rate for DeepSeek-R1 is 0.417 on MMLU versus 0.012 on GPQA-Diamond (probe vs. monitor), indicating a much larger gap on easy recall tasks.
• Inflection points (backtracks, realizations, reconsiderations) occur about twice as often in low-confidence traces (0.045 vs 0.020 per step for DeepSeek-R1 on MMLU).

Surprising findings

• Inflection-to-probe-shift co-occurrence is inconsistent: for GPT-OSS inflections tend to precede probe shifts, and for GPQA-D the trend reverses.
• Linear probes perform near chance (31.85%) while attention probes reach 87.98% on DeepSeek-R1 MMLU.

Limitations

• Results rely on a single LLM monitor (Gemini-2.5-Flash) and a small GPQA-Diamond test set (N=157).
• Probe-based early exit is only validated on MMLU-Redux with transfer to GPQA-Diamond; the relationship between inflections and belief shifts is not consistently captured by windowed co-occurrence.
• Forced answering prompts are off-policy and may underperform for smaller distilled models.

Experimental setup

Datasets: MMLU-Redux 2.0 GPQA-Diamond

Benchmarks: MMLU-Redux 2.0 GPQA-Diamond

Metrics: accuracy macro accuracy performativity rate (slope difference) tokens saved calibration

Baselines: linear probe forced answering Gemini-2.5-Flash CoT monitor random-label attention probe

Builds on: attention probes early answering

Closest papers by meaning

• Deep Hidden Cognition Facilitates Reliable Chain-of-Thought Reasoning (AAAI 2026 · similarity 0.89)
• Markovian Transformers for Informative Language Modeling (ICLR 2026 · similarity 0.89)
• How Does Chain of Thought Think? Mechanistic Interpretability of Chain-of-Thought Reasoning with Sparse Autoencoding (AAAI 2026 · similarity 0.89)
• TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning (NeurIPS 2026 · similarity 0.89)
• Reasoning Models Sometimes Output Illegible Chains of Thought (NeurIPS 2025 · similarity 0.88)
• Tracing the Traces: Latent Temporal Signals for Efficient and Accurate Reasoning (ICLR 2026 · similarity 0.88)