The paper introduces TASCO, a test-time adaptation framework that improves LLM reasoning by optimizing lightweight task-level prefixes while keeping the underlying model frozen. It combines confidence scores with local stability measures, based on the observation that high-confidence reasoning is more reliable when it remains stable under small perturbations. TASCO applies two perturbation strategies, random perturbation for distributional stability and sharpness-aware perturbation for worst-case sensitivity, to steer models toward more robust outputs without external verifiers or reward models.
