The authors introduce Epistemic Stance Flexibility Probing (ESFP), a behavioral benchmark designed to measure how well large language models distinguish between externally attributed prompts and self-attributed ones. ESFP evaluates this capability across six epistemic categories using 104 controlled items, assessing lexical self-attribution, representation-level responsiveness, stance content density, and cross-condition consistency.
- Evaluations of eight frontier models from five vendors show that epistemic flexibility is largely orthogonal to general model capability.
- A 27B open-weight model matched the performance of the strongest proprietary systems in this specific metric.
- The flagship model of one family underperformed its lightweight counterpart, and reasoning-optimized models did not consistently exhibit higher flexibility.
- Stance content density provided the strongest signal for measuring flexibility, whereas surface-level lexical markers like 'I think' can change without corresponding shifts in expressed stance.
ESFP measures a model's propensity to adapt its epistemic stance under changing attribution conditions rather than serving as a general competence measure.