BabelTele: When LLMs Speak in Code Humans Can't Read
This presentation explores a radical departure from human-readable prompts and agent communication. The paper demonstrates that large language models can compress information into dense, opaque representations—mixing symbols, abbreviations, and multilingual tokens—that other models decode perfectly while humans struggle. We examine BabelTele's compression efficiency, cross-model portability, and practical gains in multi-agent systems and memory-constrained contexts, revealing a future where model-to-model communication abandons readability for raw semantic density.Script
What if the prompts and memory snippets we carefully craft for language models are wasteful, optimized for the wrong audience? This paper shows that models can communicate in a radically compressed, symbol-heavy language the authors call BabelTele, completely opaque to humans but semantically transparent to other models.
BabelTele text scores 16.7 on the Dale-Chall readability scale, with over 80% difficult words, compared to 10.3 for natural language. Human comprehension tanks, but Gemini 3.1 Pro answers questions from BabelTele with zero accuracy loss.
The method compresses input to less than 28% of original length while retaining 99.5% semantic fidelity on downstream tasks. It uses omnilingual tokens, symbolic operators, and abbreviations to maximize information density per token, with no hallucination or semantic drift.
BabelTele compressed by one model can be interpreted by others with minimal loss. The portability matrix reveals broad transferability, though some pairs like Qwen as compressor show higher reader difficulty. Compression rates vary significantly, with Gemini producing the most aggressive, opaque outputs.
In multi-agent communication, BabelTele cuts message tokens by up to 44% with task accuracy above 99%. For memory-constrained contexts like ultra-long code repositories, compression enables models to operate beyond their nominal context limits, boosting accuracy from 55% to 62% when truncation would otherwise fail.
This work decouples human readability from model decodability, opening a path toward model-native meta-languages for agent systems and memory hierarchies. If you're curious how compression reshapes Large Language Model communication, explore more research and create your own videos at EmergentMind.com.
