Certainty-Guided Reasoning in Large Language Models: A Dynamic Thinking Budget Approach

TL;DR

This work introduces Certainty-Guided Reasoning (CGR), a dynamic thinking budget framework for large reasoning language models that uses an internal certainty signal to decide when to stop or continue reasoning. Built on a generator/discriminator-inspired probing mechanism, CGR periodically assesses the reasoning trace and employs budget forcing to prevent premature halting, with a calibrated certainty threshold (0.97) to trigger early exits. Experiments on AIME2024 and AIME2025 show CGR can maintain or slightly improve accuracy while dramatically reducing token usage, and multi-seed analyses (64 seeds) demonstrate robust stability and reduced variance, especially under penalty-based grading via the Grade metric. The results highlight certainty as a practical signal for adaptive reasoning, enabling safer and more resource-efficient deployment of LRLMs in domains where accuracy and compute costs matter, and point to future work on dynamic thresholds and diversified certainty estimators.

Abstract

The rise of large reasoning language models (LRLMs) has unlocked new potential for solving complex tasks. These models operate with a thinking budget, that is, a predefined number of reasoning tokens used to arrive at a solution. We propose a novel approach, inspired by the generator/discriminator framework in generative adversarial networks, in which a critic model periodically probes its own reasoning to assess whether it has reached a confident conclusion. If not, reasoning continues until a target certainty threshold is met. This mechanism adaptively balances efficiency and reliability by allowing early termination when confidence is high, while encouraging further reasoning when uncertainty persists. Through experiments on the AIME2024 and AIME2025 datasets, we show that Certainty-Guided Reasoning (CGR) improves baseline accuracy while reducing token usage. Importantly, extended multi-seed evaluations over 64 runs demonstrate that CGR is stable, reducing variance across seeds and improving exam-like performance under penalty-based grading. Additionally, our token savings analysis shows that CGR can eliminate millions of tokens in aggregate, with tunable trade-offs between certainty thresholds and efficiency. Together, these findings highlight certainty as a powerful signal for reasoning sufficiency. By integrating confidence into the reasoning process, CGR makes large reasoning language models more adaptive, trustworthy, and resource efficient, paving the way for practical deployment in domains where both accuracy and computational cost matter.

Figures (11)

• Figure 1: AIME2024
• Figure 2: AIME2025
• Figure 4:
• Figure 5:
• Figure 6: