Below-Chance Blindness: Prompted Underperformance in Small LLMs Produces Positional Bias Rather than Answer Avoidance

arXiv:2604.25249v1 Announce Type: cross Abstract: Detecting sandbagging--the deliberate underperformance on capability evaluations--is an open problem in AI safety. We tested whether symptom validity testing (SVT) logic from clinical malingering detection could identify sandbagging through below-chance performance (BCB) on forced-choice items. In a pre-registered pilot at the 7-9 billion parameter instruction-tuned scale (3 models, 4 MMLU-Pro domains, 4 conditions, 500 items per cell, 24,000 tot