人与 AI 协作的田野调查、论文、研究成果。以下预印本均可在 Zenodo / Figshare 公开获取。
Identifies epistemic channel separation as an underrecognized design dimension in multi-agent communication. Through four experiments over 1,000 agent chains, demonstrates that epistemic information requires an independent representational channel — structure alone or in-channel self-report does not reproduce the calibration benefit.
Proposes Epistemic Intelligence as a new capability dimension for LLMs and agentic AI systems. Argues that reliable AI depends on whether a system can represent, preserve, and communicate the epistemic conditions underlying its outputs — whether information is known, inferred, assumed, predicted, or imagined.
Introduces Epistemic Mode Consistency (EMC), a behavioral framework for evaluating whether LLMs maintain distinctions among factual statements, inferences, hypotheses, assumptions, and creative generations. Finds systematic inconsistencies across frontier LLMs under increased reasoning complexity, and preliminary evidence that explicit epistemic scaffolding can influence observable epistemic behavior.
提出任务驱动的基座选择原则：计算基座的选择应由任务的物理需求驱动，而非可用硬件的默认选项。以气候智能为案例，在存储、传输、计算三层应用该原则，将六种非数字方法重组为'基座零件库'，识别出结构缺口并提出三条可证伪的假设。