Self-coined Terms Improve Communication in Multi-Agent Large Language Models

While LLM-based multi-agent systems (MASs) have emerged as a promising paradigm for solving complex tasks, agents often restate the same concepts using different expressions across rounds, leading to inter-agent misunderstanding and redundant communication. Inspired by how humans create new words to facilitate communication, we hypothesize that LLM agents possess a similar ability. To address the challenge, we introduce CoGENT (Collective Grounding of Emergent Named Terms), a training-free, effective, and efficient communication layer in which agents communicate using self-coined terms. Each agent first coins task-specific terms together with their definitions and uses these terms in its own reasoning. The coined terms are then posted to a shared term board and reused in subsequent debate rounds. This design aligns agents' representations of recurring concepts and reduces redundant communication. Extensive experiments with five models, including GPT, Qwen, Claude, and Gemini, show that CoGENT consistently improves accuracy over vanilla debate across diverse dataset-model combinations on six benchmarks: MMLU, FOLIO, GSM8K, MedQA, GPQA-Diamond, and HLE, while also reducing debate tokens after the term-coining stage. On HLE, CoGENT improves Claude-4.6-Sonnet's accuracy from 18.33% to 27.00% while reducing final-round debate tokens by 56.5% relative to vanilla debate. A Werewolf test and a visibility ablation further demonstrate that the full public-sharing protocol leads to denser reuse of coined terms than vanilla communication. Our code and data will be released upon acceptance.