Large Language Models Do Not Always Need Readable Language

The paper presents BabelTele, a novel model-centric textual representation that allows for encoding semantic information in non-human-readable forms while maintaining high semantic fidelity. It demonstrates that LLMs can achieve 99.5% semantic fidelity with representations condensed to 27.9% of their original length, suggesting a potential reduction in context overhead for downstream tasks. This approach indicates a shift towards model-native representations, which could enhance efficiency and performance in multi-agent communication and cross-model transfer scenarios.