Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Think with Confidence

Yichao Fu, Xuewei Wang, Yuandong Tian, Jiawei Zhao

TL;DR

Self-consistency improves reasoning at substantial compute cost with diminishing returns. DeepConf leverages local, model-internal confidence signals to filter and adaptively terminate low-quality traces, achieving high accuracy with far fewer tokens in offline and online settings. Extensive experiments on open-source LLMs across hard math benchmarks show near-perfect performance on AIME 2025 and substantial token savings, demonstrating a practical approach to test-time compression for reasoning. The method is simple to deploy, requires no training, and generalizes across model scales and frameworks, offering a scalable path to efficient reasoning in production.

Abstract

Large Language Models (LLMs) have shown great potential in reasoning tasks through test-time scaling methods like self-consistency with majority voting. However, this approach often leads to diminishing returns in accuracy and high computational overhead. To address these challenges, we introduce Deep Think with Confidence (DeepConf), a simple yet powerful method that enhances both reasoning efficiency and performance at test time. DeepConf leverages model-internal confidence signals to dynamically filter out low-quality reasoning traces during or after generation. It requires no additional model training or hyperparameter tuning and can be seamlessly integrated into existing serving frameworks. We evaluate DeepConf across a variety of reasoning tasks and the latest open-source models, including Qwen 3 and GPT-OSS series. Notably, on challenging benchmarks such as AIME 2025, DeepConf@512 achieves up to 99.9% accuracy and reduces generated tokens by up to 84.7% compared to full parallel thinking.

Deep Think with Confidence

TL;DR

Self-consistency improves reasoning at substantial compute cost with diminishing returns. DeepConf leverages local, model-internal confidence signals to filter and adaptively terminate low-quality traces, achieving high accuracy with far fewer tokens in offline and online settings. Extensive experiments on open-source LLMs across hard math benchmarks show near-perfect performance on AIME 2025 and substantial token savings, demonstrating a practical approach to test-time compression for reasoning. The method is simple to deploy, requires no training, and generalizes across model scales and frameworks, offering a scalable path to efficient reasoning in production.

Abstract

Large Language Models (LLMs) have shown great potential in reasoning tasks through test-time scaling methods like self-consistency with majority voting. However, this approach often leads to diminishing returns in accuracy and high computational overhead. To address these challenges, we introduce Deep Think with Confidence (DeepConf), a simple yet powerful method that enhances both reasoning efficiency and performance at test time. DeepConf leverages model-internal confidence signals to dynamically filter out low-quality reasoning traces during or after generation. It requires no additional model training or hyperparameter tuning and can be seamlessly integrated into existing serving frameworks. We evaluate DeepConf across a variety of reasoning tasks and the latest open-source models, including Qwen 3 and GPT-OSS series. Notably, on challenging benchmarks such as AIME 2025, DeepConf@512 achieves up to 99.9% accuracy and reduces generated tokens by up to 84.7% compared to full parallel thinking.

Paper Structure

This paper contains 51 sections, 12 equations, 11 figures, 12 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Confidence as an Indicator of Reasoning Quality
  3. Token Entropy.
  4. Token Confidence.
  5. Average Trace Confidence.
  6. Deep Think with Confidence
  7. Confidence Measurements
  8. Group Confidence.
  9. Bottom 10% Group Confidence.
  10. Lowest Group Confidence.
  11. Tail Confidence.
  12. Offline Thinking with Confidence
  13. Majority Voting.
  14. Confidence-Weighted Majority Voting.
  15. Confidence Filtering.
  16. ...and 36 more sections

Figures (11)

  • Figure 1: Up: DeepConf on AIME 2025. Down: Parallel thinking using DeepConf.
  • Figure 2: Confidence distributions for correct vs. incorrect reasoning traces across different metrics. Data from HMMT25: 30 problems, 4096 traces each.
  • Figure 3: Confidence measurements and offline thinking with confidence.
  • Figure 4: DeepConf during online generation.
  • Figure 5: Offline accuracy with Lowest Group Confidence filtering (DeepSeek-8B) on AIME24, AIME25, BRUMO25, and HMMT25. The $\eta\%$ variant retains only the top $\eta\%$ highest-confidence traces before confidence-weighted majority voting.
  • ...and 6 more figures