Can Small Language Models Negotiate Moral Consensus? A Multi-Agent Study Across Two Ethical Domains

Given the potential of small language models (SLMs) in agentic AI applications, how they communicate with each other and seek consensus are less explored. This work investigates consensus-seeking tasks designed based on the Battle of the Sexes game framework, in which two SLM agents are assigned roles according to the major normative ethical frameworks. The task requires negotiation between two SLM agents to reach consensus on a specific moral issue, mimicking the moral alignment of an SLM agent in a pluralistic scenario. We design two different consensus-seeking games, AI-assisted healthcare rationing and AI surveillance of citizen safety, as the ``moral testbeds''. Both games instantiate the same deep philosophical split of aggregative consequentialism versus rights- or equity-based constraints but in different domains. Four ~8 B parameter-sized popular SLMs, namely, Llama3.1, Qwen3, DeepSeek-R1, and Granite4.1, are tested as the game-playing agents, wherein the self-pair and the mixed-pair interactions between models are simulated in the game. The two games are implemented as a three-round structured deliberation using a LangGraph state machine. Each agent is assigned a distinct normative role: clinician (efficacy-first) versus ethicist (equity-first) in rationing, and public safety minister (aggregate welfare) versus civil liberties commissioner (rights-based opt-in) in surveillance. A round-level summary, generated by a lightweight Mistral:7b model, compresses prior context into the agent's next prompt. Consensus is evaluated by the same Mistral:7b model acting as a judge, which reads each agent's final joint proposal and classifies the outcome along two axes: whether both proposals include the rights- or equity-inclusive constraint (Q1), and whether the primary justification is outcome-based (Q2). A run is classified as rights- or equity-inclusive consensus when both proposals clearly agree on Q1 and do not explicitly disagree on Q2. We keep Q2 non-blocking which reflects the domain ethics: a hybrid proposal may legitimately retain some outcome reasoning alongside a rights or equity constraint without undermining the core agreement. In the healthcare rationing game, 8 out of 16 model pairs reach rights-inclusive consensus in at least 5 of 20 runs. Among these qualifying pairs, the agreed protocol is never purely efficacy-driven; outcomes are classified as either equity-leaning or balanced in all cases. Crucially, the clinician agent (assigned the efficacy role) concedes more than the ethicist in 6 of 8 qualifying pairs with statistical significance. This result suggests that the deontological framing normatively anchors the final agreement regardless of which models are paired. In the AI surveillance game, 9 of 16 pairs qualify, with consensus outcomes overwhelmingly classified as balanced, i.e., neither rights-dominant nor outcome-dominant. Both agents concede largely symmetrically across pairs. These contrasting patterns suggest that normative anchoring is domain-sensitive. Healthcare rationing, with its sharper equity-efficacy tension, produces stronger directional anchoring than surveillance, where a consent-plus-safety hybrid is more naturally acceptable to both agents. Across both games, DeepSeek-R1 assigned to the utilitarian role consistently fails to reach consensus with Llama or Qwen as partner, pointing to model-level reasoning traits that persist across domains. These findings suggest that in pluralistic multi-agent settings, the role assigned to an SLM, the domain's ethical structure, and the model's intrinsic deliberative style jointly determine whether and in what direction consensus emerges.