We value your privacy
This site uses cookies to improve your browsing experience, analyze site traffic, and show personalized content.
DeepThink with Confidence (DeepConf) is an efficient test-time method for large language models (LLMs) that uses model-internal confidence signals to filter out low-quality reasoning traces either during generation (online) or after generation (offline), without needing any extra training or hyperparameter tuning. Incorporating local confidence metrics such as lowest-group, bottom-10%, and tail confidence, DeepConf dynamically prioritizes high-quality reasoning paths and can terminate poor traces early, reducing both token usage and computational overhead substantially.