Researcher at Unknown
Co-author of the Deep Think with Confidence research paper
How media typically covers Yichao Fu
Yichao Fu as author
Meta's DeepConf method improves LLM reasoning to 99.9% accuracy on AIME 2025 while reducing generated tokens by 84.7%, requiring only ~50 lines of code with no model retraining.
“Co-author of the DeepConf research paper on parallel thinking methods for LLM reasoning”
DeepConf, a parallel thinking method using model-internal confidence signals, achieves 99.9% accuracy on AIME 2025 while reducing generated tokens by 84.7% compared to standard thinking approaches without requiring additional training.
“Co-author of the Deep Think with Confidence research paper”