Meta & UCSD Develop DeepConf: Boosting Language Model Efficiency & Accuracy
Researchers from Meta and the University of California, San Diego, have developed DeepConf, a method that enhances language models' efficiency and accuracy. It operates in offline and online modes, with the latter stopping low-confidence paths early, leading to significant token consumption reductions.
DeepConf achieved remarkable results on the gpt-oss-120B model for AIME 2025. In offline mode, it reached 99.9% accuracy, and in online mode, it scored 97.9% with an 84.7% reduction in token consumption. This method filters out low-quality solution paths by analyzing a model's confidence in its predictions.
The aggressive variant of DeepConf reduced token consumption by up to 84.7% in mathematical tasks, maintaining accuracy. The conservative variant saved up to 59% without compromising performance. Notably, DeepConf does not require additional training and can be integrated into existing systems with minimal code changes.
DeepConf's ability to improve mathematical reasoning in language models, reduce computational cost, and increase accuracy positions it as a crucial player in the evolution of language models. However, concerns about long-term viability due to rising energy costs and the need to prove economic viability have been raised.
Read also:
- Meta's AI Chatbot Boosts Content, Apple's Siri Under French Data Probe
- Europe's Energy Infrastructure Under Siege: Cyberattacks Surge 67% in 2025
- YouTube has disbursed over $100 billion to content creators on its platform since the year 2021.
- Investment of $20 million in strategy by the Aqua 1 Foundation of the UAE in Above Food
