AI models may not need to speak in plain English — at least not when talking to each other.
Researchers have developed a framework called BabelTele that lets large language models encode and decode semantic information in compact, human-unreadable formats. The paper, posted to arXiv, tested these compressed representations across instruction-tuned models and found that text can be shrunk to 27.9% of its original size while retaining 99.5% semantic fidelity. The researchers evaluated BabelTele across agent memory tasks, cross-model transfer, and multi-agent communication — not just toy benchmarks.
Most AI systems today route everything through natural language, even in model-to-model exchanges where no human ever reads the output. That design choice burns context window space and adds latency. If models can reliably communicate in denser, non-standard formats, multi-agent pipelines could become significantly cheaper and faster to run — a real concern as those architectures scale.
The caveat the paper buries but shouldn't: performance depends heavily on the specific compressor-reader model pair, meaning BabelTele is not a plug-and-play protocol yet. It's closer to a proof of concept than a shipping standard — but it's a proof of concept that chips away at one of the quieter assumptions baked into how LLM systems are built today.