LLM

Reasoning

Advancing LLM Reasoning: Deep Think with Confidence // DeepConf

DeepConf: Efficient LLM Reasoning via Confidence-Based Early Stopping and Adaptive Consensus

https://arxiv.org/html/2508.15260v1

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.

1. Core Problem

The paper addresses the high computational cost of advanced reasoning methods in Large Language Models (LLMs), such as Self-Consistency and Chain-of-Thought (CoT). While these methods improve accuracy by generating multiple reasoning traces, they are inefficient, often producing many tokens for low-quality or incorrect reasoning paths.

2. Key Insight

Not all reasoning traces are equally valuable. The quality of a reasoning trace can be estimated using confidence metrics derived from the model’s internal token distributions. By identifying and filtering out low-confidence traces early, computational resources can be saved without sacrificing accuracy — or even improving it.

3. Proposed Method: DeepConf

DeepConf is a framework that leverages confidence signals to make LLM reasoning more efficient and effective. It operates in two modes:

a) Offline Thinking with Confidence

• After generating a fixed set of traces, DeepConf uses trace-level confidence scores to weight or filter traces before voting.
• Three confidence metrics are evaluated:
• Average Trace Confidence: Mean confidence over all tokens.
• Bottom-10% Group Confidence: Mean of the lowest 10% of group confidences in a trace.
• Tail Confidence: Confidence over the final 2,048 tokens.
• Key finding: Bottom-10% and Tail confidence better distinguish correct from incorrect traces than average confidence.

b) Online DeepConf (Main Contribution)

• Implements early stopping during trace generation.
• Uses a warm-up phase (e.g., 16 initial traces) to calibrate a confidence threshold s based on the top η% of trace confidences.
• Any new trace is terminated early if its current group confidence drops below s.
• Two variants:
• DeepConf-low (η = 10%): Aggressive filtering, high savings, risk of overconfidence.
• DeepConf-high (η = 90%): Conservative filtering, moderate savings, more robust.
• Adaptive sampling stops further trace generation once a consensus threshold (τ = 0.95) is reached.

4. Key Innovations

• Group-based confidence: Uses sliding windows (e.g., 2048 tokens) to compute local confidence, enabling detection of reasoning breakdowns.
• Lowest/Bottom-10% Group Confidence: Focuses on the weakest segment of a trace as a proxy for overall quality — more effective than averaging.
• Early stopping: Reduces wasted computation on low-quality traces.
• Adaptive consensus stopping: Dynamically decides when enough high-confidence traces have been generated.

5. Experimental Findings

• Datasets: AIME24/25, BRUMO25, HMMT25 (math competitions), GPQA-Diamond (STEM).
• Models: DeepSeek-8B, Qwen3–8B/32B, GPT-OSS-20B/120B.
• Results:
• DeepConf-low reduces token usage by 60–78% while matching or exceeding majority voting accuracy.
• Example: On AIME25 with DeepSeek-8B, achieves 86.4% accuracy with 1.24×1⁰⁸ tokens vs. 82.3% with 4.01×1⁰⁸ for majority voting.
• DeepConf-high offers more balanced savings (39–59%) with stable accuracy.
• Confidence-weighted voting outperforms unweighted voting, especially with aggressive filtering (top 10%).
• Bottom-10% confidence is a superior quality estimator compared to mean confidence.

6. Trade-offs and Limitations

• Overconfidence risk: Models can be highly confident on incorrect traces, leading to accuracy drops if filtering is too aggressive (e.g., GPT-OSS-120B on HMMT25).
• Calibration sensitivity: Performance depends on proper threshold setting via warm-up.
• Task dependency: Effectiveness varies across datasets and models.

7. Broader Implications

• Efficient test-time scaling: DeepConf enables better accuracy-cost trade-offs, crucial for deploying reasoning-heavy LLMs.
• Confidence as a control signal: Internal confidence metrics can guide inference dynamics (e.g., early stopping, adaptive depth).
• Future directions: Integration with reinforcement learning, improved calibration, and uncertainty quantification.

8. Conclusion

DeepConf demonstrates that leveraging model-intrinsic confidence signals — especially focusing on the least confident parts of reasoning — can significantly improve the efficiency and accuracy of LLM reasoning. By combining confidence-weighted aggregation with early stopping and adaptive sampling, it achieves state-of-the-art performance with far fewer computational resources than standard self-consistency methods.