Reproduce “Feeling of Knowing in Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zichuan Fu, Xian Wu, Jingtong Gao, Wenlin Zhang, Jiaxuan Li, Binhao Wang, Yimin Deng, Guojing Li and 4 more
The authors train a model to keep its answerability judgments aligned as its knowledge changes. They update knowledge, train a separate functional feeling of knowing channel across intermediate states, and train responses to follow that judgment. After one TriviaQA update, Qwen2.5-7B achieved 0.692 held-out answer accuracy and 0.809 AUROC, versus 0.546 and 0.519 for vanilla knowledge SFT; on the unupdated model, EliCal reached 0.810 AUROC.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.