🔶 前沿推理 Agent 在交互式二维迷宫中失效

英文标题：🔶 Frontier Reasoning Agents Fail on Interactive 2D Mazes

MultiNet 2.0 预览报告评估了前沿视觉语言模型（VLM）在交互式二维环境中的目标导向能力。报告指出，随着任务时间跨度（task horizon）增加，模型的目标进度会逐步衰减。该结果表明，当前前沿 Agent 在需要持续规划、状态跟踪与长期执行的交互任务中仍存在明显局限，对开发可靠的长程自主智能体具有重要参考价值。

打开原文相关 5 · 机会 1

中文

Guruprasad, P., Rivera, S., Lu, H., Jain, A., Ren, H. 和 Sikka, H.（2026）。《MultiNet 2.0 预览：前沿 VLM 在交互式二维环境中的目标进度会随任务时间跨度而衰减》。获取地址：https://www.fig.inc/blog/multinet-v2-preview

bibtex @online{multinet_v2_preview_technical_report_2026, title = {MultiNet 2.0 Preview: Goal Progress Decays with Task Horizon for Frontier VLMs in Interactive 2D Environments}, author = {Pranav Guruprasad and Sean Rivera and Helen Lu and Arushi Jain and Hangliang Ren and Harshvardhan Sikka}, year = {2026}, url = {www.fig.inc/blog/multinet-v2-preview}, note = {MultiNet 2.0 Preview: Interactive 2D Mazes} }

English / 原文

__ Copy

Guruprasad, P., Rivera, S., Lu, H., Jain, A., Ren, H. and Sikka, H. (2026) MultiNet 2.0 Preview: Goal Progress Decays with Task Horizon for Frontier VLMs in Interactive 2D Environments. Available at: https://www.fig.inc/blog/multinet-v2-preview

__ Copy

@online{multinet_v2_preview_technical_report_2026,
  title   = {MultiNet 2.0 Preview: Goal Progress Decays with Task Horizon
             for Frontier VLMs in Interactive 2D Environments},
  author  = {Pranav Guruprasad and Sean Rivera and Helen Lu and Arushi Jain and
             Hangliang Ren and Harshvardhan Sikka},
  year    = {2026},
  url     = {www.fig.inc/blog/multinet-v2-preview},
  note    = {MultiNet 2.0 Preview: Interactive 2D Mazes}
}