Apparent Local Robustness Under Context-Reduced Interfaces: A Minimal Synthetic Stress Test
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
This technical note presents a minimal synthetic stress test on local robustness under reduced context interfaces in large language model inference. A controlled dataset of 100 multiple-choice instances is introduced across four evidence regimes: easy local retrieval, distributed evidence integration, conflict-sensitive update resolution, and position-sensitive retrieval. The same model is evaluated under three interfaces: FULL, RECENT, and a deterministic COMPRESSED interface retaining only the first sentence of each block. The main result is a clear regime-dependent dissociation: perfect local robustness is preserved in the easy regime, while strong interface-specific failures emerge outside it. The note is strictly diagnostic and local in scope, and argues that local benchmark robustness under a reduced context interface does not imply global behavioral equivalence across evidence regimes.
Fields of Study & Subjects
Topic Modeling
Adversarial Robustness in Machine Learning
Natural Language Processing Techniques
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 27 March 2026
- Language
- EN