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

Does Thinking More Always Help? Mirage of Test-Time Scaling in Reasoning Models

NeurIPS 2025 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 1 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.

Shows that prompting reasoning models to keep thinking ("Wait") initially helps then hurts (overthinking), attributes early gains to rising output-entropy variance, and proposes parallel thinking with majority vote under a fixed token budget.

Problem

Whether extending test-time thinking via prompts like "Wait" reliably improves reasoning in reasoning models; the paper finds accuracy is non-monotonic in thinking length.

Approach

Empirically identifies an overthinking critical point under budget-forced "Wait" prompting, explains the early gains as a variance (entropy) effect via a simple Gaussian reward-overlap model, and proposes parallel independent reasoning traces with self-consistency-style majority vote as a better use of the fixed budget.

Key results

• Under Wait & Think More on GSM8K with DeepSeek-R1-Distill-Qwen-1.5B, accuracy rises from 82.2% to 87.3% at ~1100 thinking tokens, then falls to 70.3% at ~15980 tokens.
• Parallel thinking at a 16K token budget yields about 22% higher accuracy than Wait & Think More and about 47% higher than Exact Thinking budget control.
• On GSM8K, output-answer entropy rises from 0.23 to 2.79 when average thinking grows from 385 to 6136 tokens.

Measured effects reported in the paper

Surprising findings

• Appending repeated copies of the same reasoning step does not raise answer entropy, suggesting that overthinking rather than raw context length drives variance.
• Accuracy declines with longer thinking on GSM8K, MATH-500 and AIME 2024 across the evaluated distilled models, contrary to the common belief that more thinking always helps.

Limitations

• Only mid-sized models (1.5B to 8B) were studied due to resource constraints.
• The variance explanation is supported by empirical evidence and a simple one-dimensional Gaussian model rather than a formal theory.

Experimental setup

Datasets: GSM8K MATH-500 AIME 2024

Benchmarks: GSM8K MATH-500 AIME 2024

Metrics: accuracy output entropy

Baselines: Wait & Think More Exact Thinking budget control standard thinking

Builds on: Best-of-N sampling budget forcing self-consistency

Closest papers by meaning

• Rethinking External Slow-Thinking: From Snowball Errors to Probability of Correct Reasoning (ICML 2025 · similarity 0.91)
• Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning (NeurIPS 2025 · similarity 0.90)
• Rethinking Thinking Tokens: LLMs as Improvement Operators (ICML 2026 · similarity 0.90)
• Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models (NeurIPS 2025 · similarity 0.90)
• Longer Context, Deeper Thinking: Uncovering the Role of Long-Context Ability in Reasoning (NeurIPS 2025 · similarity 0.89)
• Tracing the Traces: Latent Temporal Signals for Efficient and Accurate Reasoning (ICLR 2026 · similarity 0.89)