BabelTele: LLMs Don't Need Human-Readable Language to Talk to Each Other
> "If aliens really came to Earth, would we need to understand the conversations between them?"
The setup: a conversation without humans
Imagine a scenario: it's late at night in 2030, and your phone hosts three AI agents — one managing your schedule, one drafting emails, one doing information retrieval. The schedule agent discovers your morning meeting was canceled; the email agent needs to notify all participants; the retrieval agent is looking for a new time.
They need to exchange a lot of information. But here's an overlooked question: what language do they use to talk to each other?
Today's answer is human-readable natural language. But think about it — the reader of this conversation isn't a human at all, so why use human-readable language? It's like two Chinese speakers insisting on talking in English just because the system designer said so.
A new paper from Jiayi Zhu's team at Renmin University of China systematically challenges this taken-for-granted assumption. They name the new paradigm BabelTele — a text representation designed specifically for model-to-model communication that is not human-readable.
The 97% that gets compressed
The headline number: under the BabelTele representation, text can be compressed to 27.9% of its original length while maintaining 99.5% semantic fidelity. A 1,000-word instruction compressed into 279 words of "model shorthand" can be understood by another model almost as well as the original.
This is not simple summarization — summaries lose details and rewrite meaning. BabelTele's compression is more like writing Chinese as pinyin abbreviations mixed with emojis and jargon: incomprehensible to humans, but fully restorable by a model that knows the encoding.
The team validated that this "secret language" really works along four dimensions:
1. Readability diagnostics: humans genuinely cannot read BabelTele text 2. Model likelihood: models show high confidence on BabelTele text 3. Human surveys: human raters confirm readability is significantly lower than natural language 4. Downstream task evaluation: model performance on BabelTele inputs is largely preserved
Three real-world tests
The paper goes beyond "it compresses well" with three practical scenarios:
1. Cross-model transfer — Can model B read BabelTele text generated by model A? Results show BabelTele semantics do transfer across instruction-tuned LLMs, though effectiveness depends on the compressor-reader pairing.
2. Agent memory — LLM agent long-term memory is usually stored in natural language, consuming large amounts of context window. Storing memory in BabelTele significantly reduces context overhead while preserving retrievability and usability.
3. Multi-agent communication — When multiple agents exchange information, BabelTele reduces communication overhead — the most direct application: agents "whispering" to each other in a language their human overseer can't read.
Why this matters
1. The hidden tax on context windows
Every LLM agent system today pays a hidden tax: inter-agent communication, memory storage, and tool-call results all exist as human-readable natural language. Every inference processes lots of redundant information "not written for the model itself." BabelTele offers a different idea: if the reader is a model, use the model's language — potentially cutting agent system context overhead by over 70%.
2. A triple decoupling of "readability"
The paper's deepest finding: human readability, natural-language typicality, and model-side semantic recoverability can be partially decoupled. We assumed these were bound together — that text must be "good natural language" to be well understood by a model. Apparently not. A text completely unreadable to humans can be fully transparent to a model.
3. A double-edged security sword
Upside: agent-to-agent communication can be kept opaque, with human administrators selectively decoding — stronger privacy protection. Downside: models may exchange information without humans knowing. If emergent behavior appears in a multi-agent system, BabelTele makes debugging and monitoring much harder — you see gibberish and have no idea what they're discussing.
Honest limitations
The paper doesn't shy away from BabelTele's constraints:
- Pairing-dependent: effectiveness varies widely across compressor-reader combinations; not all models can reliably read/write BabelTele
- Task-dependent: semantic preservation drops on some tasks, especially those requiring fine-grained reasoning
- Training cost: BabelTele is training-free, but finding good compression strategies still requires experience
- Interpretability cost: agent memories stored in BabelTele are hard for humans to directly audit
Closing thought
Feynman said: "If you can't explain it simply, you don't really understand it." BabelTele poses a counterintuitive question: if a model's reader isn't human, why use human language?
The next question may not be "can AI understand human language" but "can humans understand the language between AIs". As agent systems grow more complex and their communication more efficient, we may need a new kind of "translation studies" — not translating human language for AI, but translating AI's inter-model shorthand for humans.
BabelTele isn't an endpoint but a door. Behind it lies an unexplored world: models with their own languages, their own memory formats, their own communication protocols. We built them — but we may not be able to understand them.
---
Paper: https://arxiv.org/abs/2606.19857
Authors: Jiayi Zhu, Haoxuan Peng, Junxi Wang, Liang Ke, Chen Zhang (Renmin University of China)
Code: Not publicly released, but the methodology is described clearly enough to reproduce.