When multiple agents reason with large language models (LLMs), a central question for collective decision‑making arises: whose answer should be trusted when agents disagree? Resolving such disagreements is hard because agents’ outputs are vulnerable to non‑semantic biases such as majority pressure and response‑style preferences, which can mask the quality of the underlying reasoning. Our analysis shows that bias can be both mitigated and exploited: bias control improves the selection of the initial answer, while bias injection creates decision instability that serves as a reliability signal. Building on these insights, we propose a test‑time aggregation framework that requires neither validation data nor prior knowledge of relative model strength. The framework implements bias control by presenting a single style‑normalized reasoning trace for each distinct answer, and bias injection by placing each aggregating agent’s initially selected answer in the minority. Across three mathematical reasoning benchmarks and four heterogeneous LLM combinations, our method achieves higher average accuracy and better token efficiency than competing non‑oracle aggregation methods. When applied to multi‑agent debate and Mixture‑of‑Agents, it also improves final‑answer accuracy without altering their interaction or synthesis processes.
Review: This work demonstrates that carefully managing and even leveraging bias can substantially boost the reliability and efficiency of multi‑model aggregation without extra data, offering a practical pathway for real‑world deployments.