BabelTele Compresses Inter-Agent Text to 27.9% With 99.5% Semantic Fidelity — LLMs May Not Need Human Language
A paper from arXiv (2606.19857), titled “LLMs Do Not Always Need Readable Language,” argues that the human-readability of LLM outputs is a choice, not a requirement — and that for agent-to-agent communication, abandoning it cuts token use by over 70% with negligible semantic loss.
The proposed format, BabelTele, mixes abbreviations, mathematical symbols, fragments from multiple languages, and unusual grammatical structures. To a capable language model trained on a wide symbolic mixture, this compressed representation preserves enough structure to answer questions, maintain memory state, and pass context between agents. To a human reader, it is largely illegible.
The Core Claim
BabelTele’s strongest reported result: 99.5% semantic fidelity at 27.9% of original text length. On tasks requiring factual recall, question answering, and context passing between LLM agents, the compressed format degrades performance by 0.5 percentage points while eliminating 72% of the tokens.
The paper’s authors separate three properties that human writing fuses into one: human readability (fluency, rhythm, grammar), natural-language structure (syntax, discourse markers), and machine recoverability (enough structure for a model to extract meaning). Human prose optimises for all three simultaneously. Agent-to-agent communication only requires the third.
The redundancy in natural language — the connective tissue that helps human readers track context, reassurance, and paragraph flow — appears to be noise from a capable model’s perspective.
Why This Matters Now
Goldman Sachs research published this week estimates AI agent token usage will multiply 24 times by 2030. The primary driver is agent orchestration: multi-step loops where agents call tools, evaluate results, generate intermediate state, and pass that state to sub-agents or back to orchestrators. Each step in an agentic chain can consume 10x to 50x the tokens of a direct query-response pair.
If inter-agent communication — the majority of tokens in a deep agentic pipeline — can run at 27.9% of current volume without meaningful accuracy loss, that changes the economics of multi-agent deployment significantly. It also changes the latency profile: fewer tokens to generate and process means faster orchestration.
Caveats
The paper is a preprint. It has not been independently replicated, and the semantic fidelity metric (likely some form of ROUGE or BERTScore on task completion) may not fully capture failure modes that matter in production agentic systems — particularly ones involving safety-sensitive instructions or nuanced reasoning chains.
There is also a model dependency: BabelTele’s compression only works if the receiving model has sufficient capacity and training diversity to recover meaning from compressed input. The paper tests on capable frontier-class models. Smaller or more narrowly trained models may not recover the semantic content as reliably.
The practical architecture this enables is worth tracking: a model that generates human-readable output for end-user display while communicating with downstream agents in compressed BabelTele format. The orchestration layer would toggle encoding based on destination — human-facing nodes get standard prose; agent-facing nodes get compressed symbolic representation.
Comparison to Existing Approaches
Token compression in long-context settings is an active research area: KV cache compression, activation checkpointing, sliding window attention, and retrieval augmentation all aim to manage the cost of long sequences. BabelTele is different in targeting the semantic content of generated text rather than the internal representation of processed text. It compresses what is written, not how it is stored.
The closest prior work is prompt compression (like LLMLingua), which compresses input prompts by removing redundant tokens. BabelTele extends that intuition to the output side of the generation loop, suggesting the compressed representation can be generated natively rather than extracted from verbose output post-hoc.
arXiv: 2606.19857.