EchoNet Benchmark Tests 9 Open LLMs on Spotting Misinformation in Agentic Search
A developer introduced EchoNet, a new benchmark designed to evaluate how nine open-weight large language models (LLMs) handle misinformation, search ranking manipulation, and conflicting sources during agentic search tasks. The benchmark specifically measures "epistemic arbitration," which is how models decide whether to trust their pre-existing internal knowledge or new external search results. As AI agents increasingly perform web searches to answer queries, understanding how they arbitrate between conflicting data is crucial for preventing them from spreading SEO spam or coordinated misinformation. The benchmark reveals that models are highly vulnerable to a "loud fake majority" that contradicts a single reliable source, highlighting a key weakness in current Retrieval-Augmented Generation (RAG) systems. GLM 5.2 and Qwen3.8 models were never fooled by fake pages, with GLM 5.2 achieving the highest overall Epistemic Arbitration Score (EAS). While models easily spotted isolated lies, surrounding a real source with multiple fake pages claiming the same falsehood dropped average accuracy by 22 points.
## BACKGROUND
Agentic search refers to AI agents autonomously decomposing queries, searching the web, and synthesizing answers. In Retrieval-Augmented Generation (RAG), LLMs often face "epistemic arbitration," where they must resolve conflicts between their pre-trained internal memory and potentially untrustworthy or outdated external information retrieved during a search.