Sanxing Chen
@sanxing.bsky.social
phd-ing @duke_nlp. previously @uva_ilp @msftresearch @googledeepmind. agentic exploration & rag
RAG systems can be both shield ✅ and weapon ⚠️ in the AI misinformation arms race. Our recent paper explores this duality!
arxiv.org/abs/2410.146...
arxiv.org/abs/2410.146...
January 31, 2025 at 8:03 PM
RAG systems can be both shield ✅ and weapon ⚠️ in the AI misinformation arms race. Our recent paper explores this duality!
arxiv.org/abs/2410.146...
arxiv.org/abs/2410.146...
How does that work? Fact-checking usually need to provide evidence for the verdict, which can be instead utilized by the adversarial generator to refine their messages. 😬 We found it’s fairly easy to sample more deceitful content from a LLM with help of detector rationales.
January 31, 2025 at 8:00 PM
How does that work? Fact-checking usually need to provide evidence for the verdict, which can be instead utilized by the adversarial generator to refine their messages. 😬 We found it’s fairly easy to sample more deceitful content from a LLM with help of detector rationales.
To test AI fact-checkers on today's news, we built an "adversarial rewrite pipeline." Think of it as stress-testing for truth!
January 31, 2025 at 8:00 PM
To test AI fact-checkers on today's news, we built an "adversarial rewrite pipeline." Think of it as stress-testing for truth!
Nah, it’s solely about political claims — there are increasingly more shallow patterns that associate with fact-check labels. General fact-check sources with more diverse content are hard for outdated LLM but not RAG-based detectors.
January 31, 2025 at 8:00 PM
Nah, it’s solely about political claims — there are increasingly more shallow patterns that associate with fact-check labels. General fact-check sources with more diverse content are hard for outdated LLM but not RAG-based detectors.