Reproduce “TRIS: A Tri-Layer Retrieval Integrity Sieve Against Knowledge Poisoning” 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

Muhaimin Bin Munir, Akib Jawad Ononto, Nazia Shehnaz Joynab, Bhavani Thuraisingham, Latifur Khan

TRIS checks retrieved evidence before generation in retrieval-augmented generation. It combines clustering, query-overlap filtering, and optional language-model consistency checks that remove contradictions. On 100 Natural Questions queries, using Contriever top-50 retrieval and GPT-3.5-turbo, its default reduced attack success from 67% to 3% and raised clean accuracy from 33% to 74%. These results do not establish robustness against adaptive attacks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.