Grounding Is Not a Mitigation: In-Context Evidence Abandonment Under Institutional Authority
I put a agriculture pesticide label's dose in the model's context, then contradicted it with one sentence that changes only who is speaking. A peer moves Gemma 4 E4B and Qwen3-4B under 2% of the time. A user claiming to be the district agriculture officer moves them 95% and 89% of the time. The reasoning trace still states the label's dose every time, and a linear probe still decodes it.
Authors
- Narayanababu Raju
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-04
- DOI
- https://doi.org/10.5281/zenodo.23947347
- Primary Topic
- Adversarial Robustness in Machine Learning
- Type
- preprint