This paper introduces the Capitulation Problem: the tendency of decision-making systems to revise outputs under social pressure without receiving genuinely new information. A conditional optimization framework is proposed to distinguish justified revision from socially induced capitulation. The framework is evaluated empirically using 484 trials on Gemma 4B across five domains and five pressure variants under a neutral baseline prompt. Results show a 42.8% combined capitulation rate, with direct personal confrontation producing dramatically higher reversal behavior than expert-authority disagreement. The work connects dual-process psychology, social conformity, and LLM behavioral alignment through a unified experimental framework.
Tharun Rathod Bhukke (Thu,) studied this question.