Agent Memory Is a Surface for Endogenous Authorization Laundering
Cerruti, Okamoto, and Erol. arXiv, 2026
Privacy and Security
What models remember about their training data, how that can be measured and predicted, and what it means for the people whose data was used.
Memorization sits at the intersection of privacy, copyright, and training dynamics. Using Pythia, we showed that memorization is emergent and partly predictable from smaller models and earlier checkpoints, and later work characterizes it as several distinct phenomena rather than one. This connects directly to our research on data attribution and open-weight safety.
Selected Papers
Newest first.
Matrix-Driven Identification and Reconstruction of LLM Weight Homology
Zhang and Goldstein. NeurIPS, 2025
A Walsh Hadamard Derived Linear Vector Symbolic Architecture
Alam, Oberle, Raff, Biderman, Oates, and Holt. NeurIPS, 2024
Holographic Global Convolutional Networks for Long-Range Prediction Tasks in Malware Detection
Alam, Raff, Biderman, Oates, and Holt. AISTATS, 2024
Detecting Backdoors with Meta-Models
Langosco, Alex, Baker, Quarel, Bradley, and Krueger. Backdoors in Deep Learning @ NeurIPS, 2023
Fooling MOSS with Pretrained Language Models
Biderman and Raff. CIKM, 2022