A Paired Testing Protocol for Batch-Conditioned Refusal Robustness in LLM Serving

In three linesarXiv study on LLM refusal robustness across batch configurations. Paired testing protocol across 15 models finds 0.16% authentic safety-label flips. vLLM with BATCH_INVARIANT=1 eliminates detected instabilities (22→0 flips). Recommendation: validate refusal in actual serving environment.

Summary generated by Claude — human-verified