A new benchmark called MIRAGE has been developed to evaluate the investigative capabilities of autonomous web agents. The benchmark consists of 750 multi-step decision tasks across Wikipedia Forensics, Shopping Admin, and Reddit Moderation, each featuring a visible surface context that can be misleading and a hidden context requiring active investigation. Evaluations of eight LLM agents revealed that while agents can access relevant pages, they struggle to extract decisive evidence, often fail when faced with contradictory information, and exhibit a significant rate of investigative hallucination. AI
IMPACT Highlights a critical gap in current AI agent capabilities, suggesting a need for improved evidence retrieval and reasoning under uncertainty.
RANK_REASON The cluster contains an academic paper introducing a new benchmark for evaluating AI agents. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →