This study employs LLM-based multi-agent simulation to investigate how fairness-related behavior evolves in repeated resource allocation. Thirty agents with five idealized value orientations participate in a survival-production cycle under equal, needs-based, contribution-based, and negotiation-based allocation mechanisms. Results show that negotiation achieves the highest productive return and subjective satisfaction with the strongest consensus in our main setting, and its subjective advantage holds across eight LLMs, suggesting that deliberative allocation can help navigate the fairness–efficiency tradeoff. Agents also exhibit systematic misalignment between subjective satisfaction and objective resource outcomes, shaped by value orientations and interaction conditions. Over repeated interaction, agents’ reflection-derived fairness positions shift away from their initial orientations toward more hybrid combinations of equality-, merit-, and needs-related reasoning. To provide human-grounded structural comparison, we further benchmark GPT-4o against published human choice distributions in 16 controlled allocation scenarios, finding moderate rank-level correspondence but substantially more concentrated LLM choices. These findings characterize value-conditioned reasoning and semantic change in LLM societies while remaining computational hypotheses rather than direct evidence of human fairness behavior.

Consensus formation is pivotal in multi-agent systems (MAS), balancing collective coherence with individual diversity. Conventional LLM-based MAS primarily rely on explicit coordination, e.g., prompts or voting, risking premature homogenization. We argue that implicit consensus, where agents exchange information yet independently form decisions via in-context learning, can be more effective in dynamic environments that require long-horizon adaptability. By retaining partial diversity, systems can better explore novel strategies and cope with external shocks. We formalize a consensus-diversity tradeoff, showing conditions where implicit methods outperform explicit ones. Experiments on three scenarios – Dynamic Disaster Response, Information Spread and Manipulation, and Dynamic Public-Goods Provision – confirm partial deviation from group norms boosts exploration, robustness, and performance. We highlight emergent coordination via in-context learning, underscoring the value of preserving diversity for resilient decision-making.

Large Language Models (LLMs) have increasingly been utilized in social simulations, where they are often guided by carefully crafted instructions to stably exhibit human-like behaviors during simulations. Nevertheless, we doubt the necessity of shaping agents’ behaviors for accurate social simulations. Instead, this paper emphasizes the importance of spontaneous phenomena, wherein agents deeply engage in contexts and make adaptive decisions without explicit directions. We explored spontaneous cooperation across three competitive scenarios and successfully simulated the gradual emergence of cooperation, findings that align closely with human behavioral data. This approach not only aids the computational social science community in bridging the gap between simulations and real-world dynamics but also offers the AI community a novel method to assess LLMs’ capability of deliberate reasoning.Our source code is available at https://github.com/wuzengqing001225/SABM_ShallWeTeamUp