Reproduce “PRISM: Transfer and Limits of Activation Interventions in Tool-Calling 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
Giovana Kerche Bonás, Thales Sales Almeida, Marcelo da Silva Reis, Rodrigo Nogueira, Marcos Meideiros Raimundo
PRISM tests whether internal activation shifts change harmful tool-call proposals and transfer across languages and model sizes. It uses unsafe-response labels from BF16 models or NF4 copies to weight BF16 activations, then applies a shift without changing model weights. A five-language direction reduced judged violations by 21.95 percentage points on 11 Arabic scenarios, not establishing general multilingual safety.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.