VibeLifeBench: Benchmarking Proactive LLM Agents in Long-Horizon Tasks

Researchers have introduced VibeLifeBench, a new benchmark designed to evaluate the proactive and persistent capabilities of large language model (LLM) agents in long-horizon, real-world-like tasks. Unlike existing evaluations that rely on short, self-contained requests in static environments, VibeLifeBench simulates a dynamic world where tasks span weeks and the environment changes without prompting. The benchmark comprises 200 tasks across ten everyday-life domains, each scripted as a multi-week timeline within a simulated world of 22 mock services. The key innovation is that the world advances autonomously, requiring agents to decide when to act, ask, or remain silent, and to notice unannounced changes while maintaining a coherent plan over time. This addresses a critical gap in current AI evaluation, as everyday life assistance demands persistence and proactivity that static benchmarks fail to measure. The work is detailed in a paper on arXiv (arXiv:2608.10875), with the announcement type 'cross'. The authors argue that an agent merely responding to immediate requests will fail in such tasks, emphasizing the need for agents that can operate independently over extended periods. VibeLifeBench aims to provide a more realistic assessment of LLM agents' capabilities, potentially guiding future developments in personal AI assistants.