LLMs Develop Symbolic Languages for Efficient Multi-Agent Reasoning

While Chain-of-Thought (CoT) prompting improves large language model (LLM) reasoning, it often generates lengthy natural language rationales that are inefficient for machine processing. To address this, a new framework called Communicative Language Symbolism Routing (CLSR) allows multiple LLM agents to autonomously create, evolve, and share compact symbolic languages, known as Language Symbolism Frameworks (LSFs). CLSR employs a latent-free router that adaptively selects and combines these LSFs for each query, optimizing for both accuracy and token cost. Unlike simple prompt optimization, CLSR treats LSFs as reusable symbolic protocols with defined usage rules, which are refined through an evolutionary loop. This approach drastically reduces generated token completion latency by 3-6 times compared to standard CoT, while preserving accuracy on challenging benchmarks, and offers a more efficient paradigm for multi-agent reasoning.