Reproduce “Latent-Space Intervention for Cross-Lingual Factual Consistency: Consistency Improvements without Accuracy Drops” 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
Faeze Ghorbanpour, Constanza Fierro, Alexander Fraser, Anders Søgaard
The authors map multilingual text representations into a shared latent space and adjust them at inference time without changing base-model weights. On Aya-expanse-8B KLAR open-ended QA, AE-only raises Rank Corr from 6.68 to 7.77 while accuracy changes from 56.77% to 56.40%; mean-shift raises Rank Corr to 24.91 but lowers accuracy to 53.45%. Results cover five languages and three 8B models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.