Title: Enhancing Diversity and Identity Representation in Public Opinion Simulations via Parametric Injection

Abstract:

Large language models (LLMs) are increasingly utilized as synthetic agents to simulate public opinion, presenting a scalable and efficient alternative to traditional, resource-intensive human surveys. However, existing LLM-driven simulation approaches struggle to reflect social heterogeneity, often resulting in flattened inter-group distinctions and uniformly homogeneous responses across various demographics. This study identifies the root cause as the "Diversity Collapse" phenomenon within LLM hidden representations, a condition where distinct social identities become progressively indistinguishable as data passes through deeper network layers.

To address this challenge, we introduce Parametric Social Identity Injection (PSII), a versatile framework designed to embed explicit, parametric representations of demographic traits and value orientations directly into the intermediate hidden states of LLMs. In contrast to prompt-based persona conditioning, PSII allows for precise and controllable modulation of identity at the representation level. Our extensive experiments, conducted on the World Values Survey using various open-source LLMs, demonstrate that PSII substantially boosts both distributional fidelity and diversity. The method effectively reduces the Kullback-Leibler (KL) divergence between simulated outputs and real-world survey data while significantly enhancing overall diversity. This research offers novel insights into the representation-level control of LLM agents, marking a significant step forward in developing scalable, diversity-aware methods for public opinion simulation.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC