New Method DeepConf Boosts Math Reasoning in Language Models
Researchers have developed a new method, DeepConf, which enhances mathematical reasoning in language models. This method reduces computational costs and boosts accuracy, operating in both offline and online modes.
DeepConf works by analyzing a model's confidence in its predictions. In offline mode, it generates all reasoning paths and then filters out low-quality ones. In online mode, it evaluates quality during generation, stopping a solution path if its confidence value drops below a threshold.
The method, developed by Yu Chen, Manya Ghobadi, Niklas Beutner, and Mohammad Alizadeh, requires no additional training and can be integrated into existing systems with minimal code changes. It has been tested with five open-source models, including the large gpt-oss-120B.
DeepConf shows weaknesses when a model is overly confident in wrong answers. The researchers recommend using the conservative variant for more stable results. In tests, DeepConf achieved an accuracy of 99.9% in offline mode and 97.9% in online mode on gpt-oss-120B for AIME 2025. It also reduced token consumption by 84.7% in online mode compared to standard majority voting.
DeepConf, a new method for improving mathematical reasoning in language models, has shown promising results. It reduces computational costs, increases accuracy, and operates efficiently in both offline and online modes. Further testing and refinement are expected to address its limitations and enhance its performance.
Read also:
- Web3 social arcade extends Pixelverse's tap-to-earn feature beyond Telegram to Base and Farcaster platforms.
- Over 5,600 Road Safety Violations Caught in Manchester Trial
- Trump praises the robustness of US-UK relations during his visit with Starmer at Chequers, showcasing the strong bond between the two nations.
- Navigating the Path to Tech Product Success: Expert Insights from Delasport, a Trailblazer in the Tech Industry
