When LLMs Develop Languages: Symbolic Communication for Efficient Multi-Agent Reasoning
Abstract Overview
This paper introduces Communicative Language Symbolism Routing (CLSR), a test-time framework in which multiple LLM agents invent, evolve, and reuse compact symbolic communication protocols called Language Symbolism Frameworks (LSFs) for reasoning. Rather than relying on long natural-language chain-of-thought traces, CLSR treats these LSFs as reusable symbolic protocols with symbol inventories, usage rules, and message-passing contracts, and uses an LLM-based router to select, ensemble, or compose them per query. The method is evaluated on seven reasoning benchmarks spanning knowledge QA, multi-hop QA, and mathematics, using several open-source backbones including Qwen, LLaMA3, and DeepSeek-R1 variants. The paper also provides a theoretical framing of reasoning efficiency, including a token-accuracy optimization view and an information-theoretic lower bound on token cost under symbolic communication.
Novelty
The distinctive contribution is allowing LLMs to autonomously generate and iteratively evolve reusable symbolic languages for reasoning, then adaptive routing among those languages at inference time. This differs from prompt optimization over surface instructions and program-execution pipelines with human-designed intermediate languages, and it is paired with a formal analysis of the accuracy-token trade-off.
Results
Across multiple reasoning benchmarks, CLSR reduces latency-oriented generated completion tokens by roughly 3-6x compared to standard chain-of-thought while maintaining accuracy. The experiments also demonstrate an improved accuracy-token Pareto frontier across multiple open-source backbones, with ablations confirming the benefits of deeper language evolution and larger agent populations.
Key Points
- CLSR frames efficient reasoning as adaptive test-time routing over evolved symbolic communication protocols rather than natural-language prompts.
- The inference system dynamically chooses between a single low-cost LSF call, multi-LSF aggregation, and multi-round LSF composition based on query difficulty.
- Both empirical and theoretical analyses demonstrate that structured symbolic traces improve reasoning efficiency by increasing the task-relevant information carried per generated token.