Large Language Models Do Not Always Need Readable Language

As LLMs are increasingly used for long-context tasks, agent memory, and multi-agent communication, the reliance on human-readable natural language creates a significant bottleneck. This paper investigates whether semantic information can be encoded in compact, non-standard textual forms—termed BabelTele—that sacrifice human readability to optimize for model-side semantic density and decodability.

The authors propose a paradigm where the human-readability constraint is relaxed. Instead of standard prose, BabelTele uses a combination of multilingual lexical selection, symbolic collapse (using emojis, operators, and punctuation), and non-standard syntax. The researchers evaluate this approach across various benchmarks, including document QA (QuALITY) and meeting records (MeetingBank), using a range of proprietary and open-weight LLMs. They measure performance through semantic fidelity, compression ratios, and the impact on the reader model's reasoning process (chain-of-thought tokens).

BabelTele demonstrates that human readability and model-side semantic recoverability are partially decoupled. The method achieves high information density, maintaining 99.5% semantic fidelity even when text volume is reduced to 27.9% of its original length. Crucially, these compressed representations are not just model-specific artifacts; they exhibit robust cross-model transferability, meaning a representation generated by one model can often be interpreted by another without fine-tuning. While stronger compression can lead to longer chain-of-thought sequences as the reader model works to resolve missing information, the overall efficiency gains in context management remain significant.

This research challenges the assumption that LLM interfaces must remain human-centric. By demonstrating that models can communicate effectively using non-standard, highly compressed representations, the authors open a path toward model-native protocols that could drastically reduce context overhead in complex, long-running agentic systems and multi-agent workflows.