Distill to Detect: Exposing Stealth Biases in LLMs through Cartridge Distillation

Shayan Talaei; Abhinav Chinta; Devvrit Khatri; Amin Karbasi; Azalia Mirhoseini; Amin Saberi

Distill to Detect: Exposing Stealth Biases in LLMs through Cartridge Distillation

Shayan Talaei* Stanford University

Abhinav Chinta* Stanford University

Devvrit Khatri University of Texas at Austin

Amin Karbasi Stanford University, Cisco Systems

Azalia Mirhoseini Stanford University

Amin Saberi Stanford University

Conference on Language Modeling, 2026

How can you tell if a deployed model has been quietly taught to favor a brand or viewpoint? D2D distills the gap between a suspected model and its base into a small prefix cartridge that amplifies hidden biases into plain text.

Materials

Bibtex

@misc{talaei2026distilldetectexposingstealth,
      title={Distill to Detect: Exposing Stealth Biases in LLMs through Cartridge Distillation},
      author={Shayan Talaei and Abhinav Chinta and Devvrit Khatri and Amin Karbasi and Azalia Mirhoseini and Amin Saberi},
      year={2026},
      eprint={2607.01208},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2607.01208},
}

Scaling Intelligence Lab

Distill to Detect: Exposing Stealth Biases in LLMs through Cartridge Distillation

Materials

Bibtex