ATRBench

ATRBench is a benchmark that evaluates the ability of long-lived LLM agents to proactively ask for reusable user preferences that are not immediately needed, by fixing each user's preferences as hidden ground truth.

Papers using ATRBench (3)

AgentDoG: A Diagnostic Guardrail Framework for AI Agent Safety and Security2026Ask Now, Use Later: Benchmarking the Proactivity Gap in Long-Lived LLM Agents2026Content-Aware Attack Detection in LLM Agent Tool-Call Traffic: An Empirical Study of Features, Architectures, and Evaluation Protocols2026