arXiv:2508.06482v2 Announce Type: replace-cross Abstract: Humans communicate with increasing efficiency in multi-turn interactions, by adapting their language and forming ad-hoc conventions. In contrast, prior work shows that LLMs do not naturally show this behavior. We develop a post-training process to develop this ability through targeted fine-tuning on heuristically identified demonstrations of convention formation. We evaluate with two new benchmarks focused on this capability. First, we design a focused, cognitively-motivated interaction benchmark that consistently elicits strong convent
The paper addresses a known limitation in Large Language Models (LLMs) regarding natural, efficient multi-turn communication, a critical area for their deployment in complex interactions.
This research suggests a pathway to overcome fundamental communication inefficiencies in LLMs, which could open new frontiers for autonomous AI agents and sophisticated human-AI collaboration.
LLMs may evolve from static knowledge bases to adaptive communicators, forming ad-hoc conventions that significantly improve interaction quality and reduce communicative friction.
- · AI developers
- · Customer service industries
- · Collaboration software providers
- · AI agent platforms
- · Inefficient AI systems
- · Developers relying solely on brute-force scaling for communication improvements
LLMs will become significantly more effective in multi-turn dialogues and collaborative tasks.
Improved LLM communication could accelerate the development and deployment of truly autonomous AI agents by enhancing their ability to coordinate and adapt.
More efficient AI communication might lead to unexpected emergent behaviors in complex multi-agent systems, potentially altering work paradigms and inter-agent dynamics.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.LG