Oratis (Wang Bihao) · Diogenes · Draft v3, July 2026 · PDF
Language-model agents drift. Building on Anthropic's finding that LLMs maintain a verbalizable global workspace (Gurnee et al., 2026), this paper models persistent self-state agent architectures (LISA's soul) as functional externalizations of the workspace — and backs the model with experiments that run entirely on open models and… See the full description on the dataset page:
https://huggingface.co/datasets/DiogenesLab/externalizing-the-workspace.
