The Virtues of Brevity: Avoid Overthinking in Parallel Test-Time Reasoning

TL;DR

The paper tackles the challenge of improving test-time reasoning efficiency for large language models by using parallel sampling and selective scoring. It introduces a simple yet effective shortest-solution heuristic, arguing that models operate in two regimes: a concise conventional regime and a verbose overthinking regime, and that choosing the shortest solution biases sampling toward the conventional regime. Through empirical evaluation on DeepSeek-R1, Grok-3-mini, and Qwen3-32B across AIME and LiveCodeBench, the approach matches or exceeds the performance of self-consistency while substantially reducing token usage. Analyses of uncertainty markers and regime-shift indicators (e.g., token-length distributions, embedding distances) provide evidence for the two-regime picture and the mechanism behind the efficiency. The findings propose a practical, cost-saving alternative for parallel test-time reasoning that also applies when outputs are not directly comparable.

Abstract

Reasoning models represent a significant advance in LLM capabilities, particularly for complex reasoning tasks such as mathematics and coding. Previous studies confirm that parallel test-time compute-sampling multiple solutions and selecting the best one-can further enhance the predictive performance of LLMs. However, strategies in this area often require complex scoring, thus increasing computational cost and complexity. In this work, we demonstrate that the simple and counterintuitive heuristic of selecting the shortest solution is highly effective. We posit that the observed effectiveness stems from models operating in two distinct regimes: a concise, confident conventional regime and a verbose overthinking regime characterized by uncertainty, and we show evidence of a critical point where the overthinking regime begins to be significant. By selecting the shortest answer, the heuristic preferentially samples from the conventional regime. We confirm that this approach is competitive with more complex methods such as self-consistency across two challenging benchmarks while significantly reducing computational overhead. The shortest-answer heuristic provides a Pareto improvement over self-consistency and applies even to tasks where output equality is not well defined.

Figures (6)

• Figure 1: Pareto curve of accuracy against token usage for DeepSeek-R1, comparing efficacy of self-consistency and picking the shortest solution on 400 AIME questions (Top-left is optimal). Picking the shortest solution is more token-efficient due to early stopping; once the first solution completes, we terminate all others, saving tokens. See Appendix \ref{['app:pareto_curves']} for other models.
• Figure 2: Analysis of different trend breaks for DeepSeek-R1 on the AIME benchmark before and after the critical point, which indicates the separation between the conventional and overthinking regimes. See Appendix \ref{['app:critical_point']} for other models.
• Figure 3: Pareto curves comparing accuracy versus token usage for Grok-3-mini and Qwen3-32B on AIME.
• Figure 4: Token-usage and critical point analyses for Grok-3-mini on AIME and LiveCodeBench.
• Figure 5: Token-usage and critical point analyses for Qwen3-32B on AIME and LiveCodeBench.