Reproduce “Evaluating the Ripple Effects of Knowledge Editing in Language Models” 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
Roi Cohen, Eden Biran, Ori Yoran, Amir Globerson, Mor Geva
This research defines and evaluates "ripple effects" in knowledge editing, where updating one fact in a language model should consistently propagate to related facts. A new benchmark, RIPPLEEDITS, reveals that current parametric editing methods often fail to integrate changes deeply, while a simple in-context editing approach demonstrates superior consistency in propagating knowledge updates.
Key insights
- Current parametric knowledge editing methods exhibit poor performance in handling ripple effects, indicating that their modifications are often superficial and do not deeply integrate new information.
- A simple, non-parametric in-context editing (ICE) baseline consistently outperforms parametric methods, suggesting that LMs' inherent compositional reasoning and in-context learning can be effective for consistent knowledge updates.
- The ability of LMs to handle ripple effects varies significantly across different logical and compositional relations, with subject aliasing generally well-handled but other forms of propagation remaining challenging.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.