Reproduce “TraceRouter++: Causal Semantic Routing Realignment via Optimal Transport” 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
Chuancheng Shi, Shangze Li, Wenhua Wu, Hengzhi Chen, Fei Shen, Li Zhuo, Zhiyong Wang, Tat-Seng Chua
TraceRouter++ identifies harmful-input activation routes and softly realigns selected coordinates toward safe-reference distributions using coordinate-wise optimal transport. On Mistral-7B-Instruct, average DSR across five evaluated jailbreak settings rises from 6.6% to 98.8% ± 0.3%. However, adaptive GCG reaches 52.4% ± 1.1%, and marginal alignment does not guarantee joint-distribution alignment in these evaluations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.