LLM 中“可靠推理”的可证伪拓扑特征
1 分•作者: Keeper123•5 天前
我们预先注册了一个可测试的协议，用于检测大型语言模型（LLM）中几何一致性推理。预测在元认知状态下，消融操作会导致贝蒂数、KS熵和脆弱性增加。欢迎复现。
OSF 协议：https://osf.io/2r6v8
预印本和案例研究：https://doi.org/10.5281/zenodo/18346699
这不是哲学——这是一个可证伪的数学假设。如果你有GPU访问权限和TDA/DL专业知识，你可以在一个周末内验证它。
查看原文
We've pre-registered a testable protocol for detecting geometrically consistent reasoning in LLMs. Predicts increased Betti numbers, KS entropy, and fragility under ablation during meta-cognitive states. Open for replication.<p>OSF Protocol: https://osf.io/2r6v8
Preprint & case studies: https://doi.org/10.5281/zenodo/18346699<p>This isn't philosophy — it's a falsifiable math hypothesis. If you have GPU access and TDA/DL expertise, you can verify it in a weekend.