Towards Budget-Efficient Socially Intelligent Agents

Large language models can perform sophisticated social interactions, but they often use unnecessarily long reasoning processes, leading to high inference cost and inefficient multi-turn dialogue. We study the problem of training socially intelligent agents that can achieve target outcomes under a limited reasoning and communication budget. We propose BudgetSI, a framework that combines turn-level contribution estimation, adaptive reasoning-budget allocation, and on-policy distillation. Given a dialogue prefix, the framework estimates the minimum budget required for the current turn, compares teacher behaviors under different budget conditions, and distills budget-efficient responses into the student model. To reduce harmful compression, candidate responses are filtered using goal-progress and semantic-consistency criteria. The resulting training procedure jointly considers social goal attainment, response quality, reasoning length, and interaction cost. We evaluate the approach on goal-oriented social dialogue tasks involving negotiation, cooperation, preference conflict, and interpersonal constraints. Our study aims to show that adaptive budget allocation can substantially reduce reasoning cost while preserving or improving the final social outcome, providing a practical approach to efficient post-training of socially intelligent language agents.