---
id: 20261101-T0-06
title: "SciGame基准：让AI代理在真实和反事实世界里玩科学发现游戏"
title_en: "SciGame Benchmark Has AI Agents Play Scientific Discovery Games in Real and Counterfactual Worlds"
url: https://ai.daily.yangsir.net/daily/20261101-T0-06
issue_date: 2026-11-01
publish_date: 2026-10-31T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2610.37919
---

# SciGame基准：让AI代理在真实和反事实世界里玩科学发现游戏

arXiv新论文提出SciGame，用游戏化任务评估大模型代理的科学发现能力。作者指出既有基准的三个问题：把规律推断和编程混在一起、评分看的是数据拟合而不是对未见输入的预测、没有把经验证据和预训练知识分开。SciGame在真实世界和反事实世界两种设定下让代理做发现任务，可以区分代理是靠推理还是靠记忆。对做AI for Science的团队来说，这提供了一种更干净的评测方式，能看出模型在新规律上的泛化能力。

## English Version

**SciGame Benchmark Has AI Agents Play Scientific Discovery Games in Real and Counterfactual Worlds**

A new arXiv paper introduces SciGame, a game-based benchmark for evaluating LLM agents on scientific discovery. The authors identify three flaws in existing benchmarks: they mix law inference with programming, score data fitting instead of prediction on unseen inputs, and fail to separate empirical evidence from pretrained knowledge. SciGame has agents perform discovery tasks in both real and counterfactual worlds, distinguishing reasoning from memorization. For AI-for-science teams, it offers a cleaner evaluation that reveals generalization to new laws.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2610.37919)

**详情页**：https://ai.daily.yangsir.net/daily/20261101-T0-06

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*