Large Language Models Do Not Always Need Readable Language

June 19, 2026 | Source: arXiv AI

Tags: LLM compression, multi-agent systems, prompt engineering, context efficiency, token optimization

BabelTele demonstrates that LLMs can compress inter-model text to 27.9% of original length while maintaining 99.5% semantic fidelity, suggesting human-readable natural language is unnecessary for LLM-to-LLM communication and opening a path to model-native representations that reduce context overhead.

Details

Standard practice in multi-agent systems is to pass messages between LLMs as human-readable natural language, even when no human will ever read them. This paper investigates whether that convention is actually necessary by introducing BabelTele — a class of compact, non-human-readable textual representations that instruction-tuned LLMs can both generate and interpret.\n\nThe key empirical result: BabelTele achieves 99.5% semantic fidelity at 27.9% of the original text volume — a roughly 3.6x compression factor. Testing covers cross-model transfer, agent memory, and multi-agent communication scenarios. The method generally maintains reliable downstream task performance, though effectiveness varies with the compressor-reader model pair and the task type.\n\nThe authors show that human readability, natural-language typicality, and model-side semantic recoverability can be partially decoupled. This opens a design space for model-native intermediate representations — prompts and messages optimized for model consumption rather than human inspection. The practical tradeoff is real: debugging and auditing become significantly harder when inter-agent communication is not human-readable.