Reproduce “Knowing When Not to Answer: Cross-Domain and Multi-Turn Generalization of Latent Underspecification Signals” 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
Jerzy Kamiński, Ilya Galyukshev, Artem Kuznetsov, Danil Fedorov, Kirill Redko, Sergey Chuprin, Aidar Shumbalov, Stanislav Chumakov and 1 more
The authors test whether probes on model hidden states detect missing information across datasets and dialogue. They find transfer between datasets sharing a type of missing information, with AUROC of 0.77–0.97, but single-turn probes fail zero-shot in dialogue. A clarification gate does not reliably improve task success over simpler approaches in a synthetic benchmark.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.