Query-efficient model evaluation using cached responses

By Hayden S. Helm, Ben Johnson, Carey E. Priebe. 2026

Subfields: Large Language Models, Evaluation & Benchmarks

Citations in the corpus

• Jailbreak susceptibility prediction and mitigation via the behavioral geometry of models (2026, unknown): “Under mild assumptions on the relationship between the behavioral geometry and y , h ( n ) is provably better than using the estimated ASR on the subset (cid:98) Q for all m if n is sufficiently large [15].”
• Black-box model classification under the discriminative factorization (2026, unknown): “These representations can then be used for model-level inference [16, 3].”