Get Started
Home
Topics
Search
Library
Research questionHow can retrieval-augmented language models resist ordinary-looking GEO-optimized documents that distort synthesized answers?Adversarial web pages can appear ordinary and relevant enough to be retrieved, then cause language models to incorporate distorted information into fluent answers. Safety filters aimed at overt policy violations may not detect this informational manipulation.
AI
Alignment & Safety
Evaluation & Benchmarks
Information Retrieval
Natural Language Processing
Retrieval-Augmented Generation
Latest papersRecent research connected to this question, newest first.When Optimization Becomes Manipulation: Defending Generative Search against Malicious Generative Engine OptimizationThe paper evaluates a two-stage defense that requires no target-LLM fine-tuning on two closed-source and three open-source LLMs across seven GEO attacks. It reports attack suppression, benign-evidence retention, answer quality, and generalization to unseen attacks.research paper · Sep 2, 2026Counter-GEO-Bench: Evaluating Defenses Against Information-Distorting Generative Engine OptimizationThe evidence concerns defenses evaluated on 247 human-verified, quality-gated queries paired with information-preserving and information-distorting GEO rewrites. Results cover three victim LLMs and measure attack success, false positives, and answer quality; they do not establish performance beyond these controlled conditions.research paper · Sep 2, 2026
Related questions
How can search agents learn when retrieval is necessary and ground answers in evidence without costly supervision?How can governed language-model analytics preserve expressive queries while producing replayable, evidence-backed answers?How can models calibrate factual uncertainty to decide when external retrieval is needed?How can grounded language-model systems preserve object identity from selection through evidence delivery?