Reproduce “Edit Knowledge, Not Just Facts via Multi-Step Reasoning over Background Stories” 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
Ya Gao, Kalle Kujanpää, Pekka Marttinen, Harri Valpola, Alexander Ilin
Models are trained to use updates as intermediate steps in multi-hop questions, not merely answer direct questions about edited facts. Contextual stories link updates to prior facts; a teacher that sees the story is distilled into a student that does not. On Qwen3-32B, answer-only Multi-Hop training scores 89.0 ± 0.3 on MQuAKE-CF without CoT, versus 38.3 ± 0.7 for Single-Hop; counterfactual updates with CoT remain a limitation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.