UniClawBench is a capability-driven benchmark for proactive agents operating on dynamic, real-world tasks. It organizes evaluation around five capabilities: Skill Usage, Exploration, Long-Context Reasoning, Multimodal Understanding, and Cross-Platform Coordination. The benchmark contains 400 bilingual tasks executed in live Docker containers and scored with fine-grained, step-level checkpoints instead of static answers. Its closed-loop setup uses an executor agent, a hidden supervisor agent, and a user agent to simulate multi-turn feedback while hiding grading criteria. The authors also evaluate models across multiple agent frameworks to separate base-model capabilities from framework effects.
No heat snapshots are available in the last 24 hours.