SimLife platform and SimLife-BP benchmark for long-horizon pattern understanding (arXiv:2609.19610v1)
The paper (arXiv:2609.19610v1) introduces SimLife, a scalable simulator of long-term household life providing rich visual observations, ground-truth action logs, and synthetic audio dialogues, and SimLife-BP, a benchmark for long-context pattern understanding with 106 episodes (avg. 15.49 hours, 38.57 in‑game days) and 1,439 QA pairs. Evaluation shows current models often make surface-level predictions, rely on frequency-based heuristics rather than if‑then rule reasoning, and struggle to adapt when behavioral patterns change, highlighting gaps in long-horizon reasoning for embodied agents.