Researchers have introduced Counter-GEO-Bench, a new benchmark designed to evaluate defenses against generative engine optimization (GEO) attacks. These attacks use SEO-like techniques to inject misinformation into large language models (LLMs), distorting their synthesized answers. The benchmark pairs queries with information-preserving and distorting GEO rewrites to test defense mechanisms. Existing defenses like Granite Guardian, Llama Guard 3, and NeMo Self-Check Fact-Checking showed limited effectiveness, reducing attack success rates by at most 5.7%. A proposed baseline, C-GEO Guard, demonstrated a significant reduction in attack success rate by 47.6% with minimal impact on utility. AI
IMPACT This research highlights a novel threat vector for LLMs and proposes a more effective defense, potentially improving the reliability of generative search results.
RANK_REASON The cluster contains a research paper introducing a new benchmark for evaluating defenses against a specific type of misinformation attack on LLMs.
- C-GEO Guard
- Counter-GEO-Bench
- generative engine optimization
- Granite Guardian
- Llama Guard 3
- NeMo Self-Check Fact-Checking
- arXiv
- large language models
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
