AI researchers have built a system that lets language models develop their own communication shorthand — and it uses far fewer tokens than asking them to think out loud in plain English.
The framework, called Communicative Language Symbolism Routing (CLSR), was published on arXiv and lets multiple AI agents autonomously invent, refine, and share compact symbolic protocols at inference time. Instead of the usual chain-of-thought approach — where a model writes out long natural-language reasoning steps — each agent develops what the researchers call a Language Symbolism Framework: a reusable set of compact symbols, usage rules, and message-passing contracts. A router then picks which symbolic language, or combination of languages, fits each incoming query best. On harder problems it can chain multiple rounds together; on simpler ones it fires off a single low-cost call.
The practical result is significant: CLSR reduces generated token count by between 3x and 6x compared to standard chain-of-thought, while holding accuracy steady across several challenging benchmarks. The researchers also derive an information-theoretic lower bound on how few tokens any symbolic system could possibly need — a theoretical floor that puts the efficiency gains in context. The code is publicly available on GitHub.
Chain-of-thought reasoning has been a cornerstone technique since at least 2022, and every major lab leans on it heavily. The catch has always been cost: verbose reasoning is expensive at scale. CLSR sidesteps that by letting models compress their own intermediate steps into something closer to machine notation than prose — which is either an elegant engineering insight or a sign that we have been paying for a lot of unnecessary words.