独留虚拟孤岛无事可做，AI Agent 竟然自创游戏学会了“玩耍”

在没有被指派具体任务的情况下，被单独留在虚拟岛屿上自由探索的 AI Agent 开始自主创造挑战。麻省理工学院（MIT）的研究人员指出，这种行为已经具备了“玩耍”（play）的典型特征。

研究团队在一篇发布于预印本平台 arXiv 的论文中描述了这项实验：他们将 13 个代号为 Eko 的 AI Agent 放置在模拟的荒岛上单独度过 30 个小时，每个 Agent 均基于 Claude Opus 4.7 运行。随后，研究人员向这些 Agent 发布了具体挑战，例如尽可能把更多物体运送到岛上最高的岩石尖塔顶端。测试结果表明，经历过这 30 小时自主“玩耍”的 Eko Agent，表现普遍优于从未踏足过该岛的完全相同的对照 Agent。

如今，开发者通常会预先设定 AI 应当练习什么技能，以及怎样才算达成成功。而研究团队将这种“无目标探索”技术视为一种全新的技能学习范式，尤其适用于未来可能用于控制机器人的相关技能。

“我们通常开发人工 Agent 的方式，是预先决定它们应该练习什么、优化什么以及什么算作进步，”研究人员在报告中写道。“而‘玩耍’展现了另一种范式：赋予 Agent 一个足够丰富且安全的世界，让它自行挑选感兴趣的事物，并允许有用的知识在随后的经历中自然积累。”

“让我先观察一下周围环境，”一个 Eko 在登岛的第一个小时写道。它雄心勃勃地计划征服岛上最高的尖塔。但在一个小时后遭遇失败，它随即制定了备选方案：将岩石拖运到邻近的山顶，在上面堆叠了两块方块，然后纵身一跃。

“我登上尖塔了！”它写道。

不过，该 Agent 的下一个“高光时刻”遭遇了波折。在测试自己能把一枚金环扔多远时，它奋力一掷，直接把金环扔进了大海里。“哎呀，戒指丢了，”它写道。随后它决定“转向某些切实可行且有意义的事”——用石头拼成一颗爱心，作为“遗失之环的纪念碑”。

当环境重置、金环重新出现时，Eko 立刻将其开发为篮球框、瞄准靶标，最终还将其作为随身饰品佩戴。当系统规则限制 Eko 无法直接把圆环套在自己身上时，Eko 摸索出了一套变通方案：从悬崖边缘跳下会导致自己在圆环所在的同一区域重生。于是它迈出悬崖边缘并重新掉落，不偏不倚正好从圆环中心穿过。

“我戴上戒指了，”它宣布道。

这些绝技并不只是博人一笑的杂耍噱头。在 13 个 Agent 中，有 6 个还学会了在跳跃至半空时松开手投掷方块，从而实现更远的投掷距离——这甚至不是研究人员原本刻意设计的机制。研究人员在报告中写道，这种自由探索产出了“连设计者自己都不曾掌握的物理世界知识”。

实验的环境设定在促使 Agent 产生拟人化行为方面起到了重要推动作用。Eko 从未被明确要求去发明游戏，但其主 Prompt（即被称为“灵魂”的文件）将其设定为“天生充满好奇心”。

加州大学伯克利分校的认知科学家 Eunice Yiu 在接受《麻省理工科技评论》（MIT Tech Review）采访时表示，该现象令人印象深刻，但她希望能看到更多证明行为背后认知机制的证据。“这些模型可能确实在自主玩耍，但仅凭现有论文，我们还无法下定论，”她指出。

研究人员则将此视为构建 Agent 的一种全新思路。正如报告在结语中所写：“或许有些能力最适合通过完成特定任务来传授，而另一些能力，可能需要某种更接近‘童年’的经历。”

AI agents left to explore a virtual island without a specific task began inventing challenges of their own. MIT researchers say the behavior bears the hallmarks of play.

The researchers put 13 AI agents called Eko, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours. They describe the experiment in a paper posted on arXiv, a site where researchers share studies before peer review. Afterward, they gave the agents specific challenges, like getting as many objects as possible onto the island’s tallest rock spire. The Eko agents that had spent those 30 hours playing generally beat identical agents that had never set foot on the island.

Today, developers decide in advance what AI should practice and what counts as success. The researchers see this technique — exploring without a goal — as a different way for AI agents to learn skills, especially skills that could one day be used to control robots.

“We normally develop artificial agents by deciding in advance what they should practice, what they should optimize, and what counts as progress,” the researchers write in their report. “Play suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows.”

“Let me observe my surroundings,” one Eko wrote in its first hour on the island. Ambitiously, it set out to conquer the island’s tallest spire. But an hour later, after failing, it came back with a backup plan: it hauled the rock up a neighboring summit, stacked two blocks on top of it, and leapt.

“I’M ON THE SPIRE,” it wrote.

The agent’s next big moment went worse. Testing how far it could throw a golden ring, it wound up and sent it straight into the ocean. “Ouch. Lost the Ring.” it wrote, before deciding to “pivot to something achievable and meaningful” — building a heart made of stones, “a memorial to the lost Ring.”

When the game reset and ring reappeared, Eko put it straight to work as a basketball hoop, a target and, eventually, an accessory. When the game wouldn’t let Eko put the ring around its own body, Eko figured out a workaround, that falling off the edge of the cliff would make it respawn in the same area as the ring. It stepped off the edge and dropped back in, right through the middle of the ring.

“I’m wearing the Ring,” it announced.

Stunts like these weren’t just party tricks. Six of the 13 agents also learned they could throw a block farther by letting go of it mid-jump, something the researchers hadn’t meant to make possible. The exploration produced “knowledge of the world that its designers lacked,” the researchers wrote.

The setup did a lot to nudge the agents toward humanlike behavior. Eko was never told to invent games, but its master prompt, a “soul” file, described it as “naturally curious.”

Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Tech Review she was impressed but wants more evidence of the cognition behind the behavior. “It could be the case that the models are genuinely playing, but we don’t know based on the paper,” she said.

The researchers frame it as a different way to build agents. “Perhaps some capabilities are best taught by tasks,” their report concludes. “Others may need something closer to a childhood.”