According to Perceive Beating monitoring, the ByteDance Seed Team has released the Smart Body Benchmark Test EdgeBench to study AI's performance in long-duration tasks ranging from 12 to 72 hours. After analyzing over 38,000 hours of smart body operation records, the research team found that AI's learning curve is highly predictable, with smart body performance improving as interaction time increases, fitting precisely to a log-sigmoid curve. The researchers used graph search theory to explain this pattern, suggesting that the smart body's progress expands outward on a skill map. The Seed Team has open-sourced the benchmark test framework and the initial set of 51 tasks.
In contrast, the research institution Epoch AI released the EBR-bench evaluation results on the same day, reaching a different conclusion. In testing tabletop game environments, even with provided policy guidelines, AI showed little to no learning or self-improvement behavior after repeated attempts. This divergence indicates that AI's self-learning potential is heavily influenced by task environment and feedback loop design. In scenarios requiring continuous environmental feedback, such as system-level engineering or scientific exploration, the smart body can build up a cumulative advantage. However, in relatively complex strategy discrete rule games, on-the-fly learning still remains a significant challenge.
