text-simulation

GlobeBench 提出一个包含 1,632 个任务的标准化基准，用于评估文本世界模型模拟 RL 环境状态转移的能力；评测 20 个 LLM 后发现即使是表现最好的 GPT-6 Astra 任务成功率也仅 49.57%，且 631 个任务无任何模型解决，凸显模型作为通用世界模型仍存在显著差距。

GlobeBench introduces a standardized benchmark of 1,632 tasks across five domains for evaluating text-based world models, finding that current LLMs — led by GPT-6 Astra at 49.57% — struggle badly with propagating action effects across dependent objects.

A developer built Slop-Cops, a Reddit app where five AI agent personalities debate and rate user-submitted project ideas, using Google Gemini API and Reddit's Devvit platform.