Reproduce “IWIBench: Evaluating Gradient-Free Model Driven Intra-Weight Injection” 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
Chris von Csefalvay, Pranav Doma, Achyuthan Unni Krishnan, Tamás Földi
IWIBench evaluates whether language models can patch another model’s weights without gradients while changing requested behavior and retaining correct behavior. Editors use bounded weight tools and target queries, then submit frozen patches tested on withheld inputs. On 30 selected episodes across ten task families for one Gemma 3 4B IT target, 6.1-sol medium scored 0.4601 versus GLM 5.3’s 0.0135.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.