Paper 2508.15260

Deep Think with Confidence

Published
Aug 2025
Research lab
Meta AI
Citations
151
GitHub
409 stars

01 In brief

Summary

Deep Think with Confidence (DeepConf) is a test-time method that improves LLM reasoning efficiency and accuracy by using model-internal confidence signals to filter low-quality reasoning traces.

It addresses the diminishing returns and high computational overhead of self-consistency with majority voting.

DeepConf operates in offline mode (filtering completed traces) and online mode (early stopping during generation), using local confidence metrics like group confidence, bottom-10% group confidence, lowest group confidence, and tail confidence.

It requires no additional training or hyperparameter tuning.

Evaluated on benchmarks like AIME 2024/2025, HMMT 2025, BRUMO25, and GPQA-Diamond with models including DeepSeek-8B, Qwen3-8B/32B, and GPT-OSS-20B/120B, DeepConf@512 achieved up to 99.9% accuracy on AIME 2025 (vs.

97.0% for majority voting) and reduced generated tokens by up to 84.7% compared to full parallel thinking.

The method uses confidence-weighted majority voting and confidence filtering (retaining top 10% or 90% of traces) to improve accuracy, with online variants (DeepConf-low and DeepConf-high) using adaptive sampling and consensus thresholds to balance efficiency and performance.

02 From the paper

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.