Reproduce “Write Facts the Way Models Read Them: Lifting the Two-Hop Curse” 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
Zhiyun Shi
LEGIBLE aligns an edited bridge’s representation with the base model’s representation of the new entity, while anchoring other subjects of the same relation. On 213 MQuAKE two-hop-curse questions, it reaches 0.71 B-Acc on Qwen2.5-7B-Instruct, versus 0.54 for base-model knowledge and 0.27 for CaKE. This supports a bridge bottleneck in tested settings, not a general reasoning improvement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.