Reproduce “Circe: Training-Free LLM Self-Evaluation via Component Interaction Consistency” 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
Shuyi Miao, Zheng Li, Xu Xie, Linrui Xu, Fei Shen, Xiaoyu Du, Tat-Seng Chua
Circe scores one generated response without reference answers or repeated sampling. It compares attention updates in earlier layers with attention-induced changes in later feed-forward network outputs, then combines this alignment with hidden-state geometry and token likelihood. Across five models and six benchmarks, it had highest mean ranking performance in all four task domains, though not in every model–domain pair.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.