SimLife is a scalable platform that simulates long‑term household life, offering rich visual observations, ground‑truth action logs, and synthetic audio‑enabled dialogues. Built on this platform, the SimLife‑BP benchmark evaluates long‑context pattern understanding—the ability to infer latent behavioral rules from weeks or months of everyday observations. The benchmark comprises 106 episodes averaging 15.49 hours (≈38.57 in‑game days) and includes 1,439 question‑answer pairs. Each task probes direct, counterfactual, noisy, and inverse reasoning under varying levels of rule hints. Evaluation of state‑of‑the‑art models shows they often achieve only surface‑level prediction, lack comprehensive rule comprehension, rely on frequency‑based heuristics rather than evidence‑driven if‑then reasoning, and struggle to adapt when behavioral patterns shift. These results indicate that long‑context pattern understanding remains a major bottleneck for embodied agents, while SimLife opens a broader space for studying memory, personalization, adaptation, and long‑horizon planning in everyday human‑AI interaction.
Review: The study highlights fundamental gaps in current models’ ability to learn long‑term behavioral rules, pointing to future work on memory augmentation and causal reasoning.