🗼 BabelTele Whitepaper BT-Spec v1.0
🇨🇳 简体中文 🤖 llms.txt 📥 English PDF GitHub ↗

BabelTele Whitepaper: model-native semantic communication and memory

🗼 BabelTele Whitepaper (BT-Spec v1.0)

A Model-Native High-Density Semantic Communication & Memory System

🇨🇳 简体中文版 (Chinese) · 🇬🇧 English Edition · 📖 Live Online E-Book · 📥 Download PDF Releases

Spec Version Online E-Book PDF Release Foundational Paper INL5-181717?style=flat-square&logo=github) License

"Humanity built MCP and A2A as the transport pipes of the agentic internet; BabelTele defines the silicon-native language flowing inside them."


📖 Overview

Even when both the sender and the receiver of a message are Large Language Models—across multi-agent coordination, persistent episodic memory, and long-context RAG—today's AI stacks still force models to communicate in verbose human prose, paying a 60%–75% Anthropocentric Readability Tax in syntactic padding, monolingual spelling inefficiency, and clause linearization.

The BabelTele System (BT-Spec v1.0) eliminates human syntactic redundancy in machine-to-machine communication, establishing a weight-agnostic, black-box-compatible, zero-shot cross-model semantic protocol.

Intuitive Collapse Examples

DomainVerbose Human Natural LanguageBabelTele Encoding (Zero-Shot LLM Native)Compression
Science"During photosynthesis, plants take in carbon dioxide and water. The water is oxidized, losing electrons, while the carbon dioxide is reduced, gaining electrons. This transforms them into oxygen and glucose."🌿光合: CO2+H2O ⇨(ox/red)⇨ O2↑ + Glucose. 💧-e⁻, CO2+e⁻.45 → 11 visual units
(ASCII BPE mode: 19 tok)
Strategy"Q3 revenue is projected to rise ~30% YoY; if it lands, the Berlin team ships the mobile app in October."Q3rev📈~30%YoY; ?lands⇒Berlin👥🚀📱@Oct24 → 9 tokens
(-62.5%)
Security"If the JWT header algorithm is set to 'none' or the expiration timestamp is missing, immediately reject the request with HTTP 401 except on public healthcheck endpoints."?JWT.alg=='none' | !JWT.exp => reject(401) !endpoint:/healthz33 → 13 tokens
(-60.6%)

📊 Key Empirical Benchmarks

DimensionVerified ResultBenchmark & Experimental Setting
Density vs. FidelityCondenses text to 27.9% of original volume while retaining 99.5% downstream QA fidelityarXiv:2606.19857 task-agnostic long-document comprehension
Heterogeneous Multi-AgentCuts inter-agent token overhead by 44.21% while retaining 99.7% of baseline task scoreDeepResearch Bench (Gemini 3.1 Pro $\leftrightarrow$ GPT-5.4 zero-shot)
Agent Episodic MemorySaves ~51% memory tokens at 96.48% relative accuracy (outperforming prose summaries)LoCoMo multi-session conversational memory benchmark
Context-Window ExtensionBoosts accuracy by +10.34% over truncation on 1.65M-token code repositoriesLongBench v2 Code Repo QA Long on 200K-window GLM-5.1 (62.07% → 72.41%)

🛠️ Official Flagship Super Agent Implementation

This repository includes babeltele-agent/, our flagship autonomous Super Agent deeply modded upon INL1 (INL2) (MIT License, Cordis all-plugin architecture) and powered by BT-Spec v1.0:

DirectoryRoleCore Architectural CapabilitiesQuick Launch (Official DSH Web UI + CLI)
INL0BabelTele · DSH Super Agent (Built upon DeepSeek Harness)• Native DSH Web UI + Cordis Mod: Preserves the complete DeepSeek Harness Web UI (in-page workspace browser, Settings → Models multi-provider key manager, file tree, PTY terminal, Subagents, Agent Teams, Plan & Goal modes) while mounting the default ⚡ BabelTele Turbo preset and dedicated settings section
• Composer-Embedded 3-Stage Neutralizer: <1ms zero-key offline emotional dehydrator + js-tiktoken (o200k_base) accounting + Tier-1 Hard-Anchor Lock right inside the DSH Composer
• #BT Reasoning Sidecar & SMFT Compaction: Hooks conversation.chat.reasoning.body for instant Chinese Sidecar Lens toggling, extends BasicCompactionEngine into KV-cache-preserving [SMFT-T1] checkpoints, and injects cross-file BT-Header symbol maps
• Dual-Plane Zero-Pollution Guard: tools/pre-execute ArtifactSanitizer gate guarantees 100% pure native Python / Java / TS / Go / HTML source code on disk
cd babeltele-agent
pnpm install && pnpm web

📚 Complete Whitepaper & Modular Specification

SectionDocumentScope & Highlights
CanonicalINL0Complete Single-Document Whitepaper (Classic 10-section architectural treatise)
Chapter 01Manifesto & The Readability TaxBiological prose redundancy, MCP/A2A channel limits, and the Babel + 1866 Telegraph metaphor
Chapter 02Theoretical Foundations & DecouplingRate-distortion projection, 13x–20x base-model PPL surge, and the Tripartite Decoupling Principle
Chapter 03Core Axioms & Canonical GrammarBT-Spec v1.0: Three axioms, tiktoken BPE paradox proof ($\Delta L_{\text{BPE}} < 0$), and six operator clusters
Chapter 04Compression Tiers & CoT DynamicsSpace-time CoT trade-off, denoising acceleration, lifecycle cost formula, and BT-L0–BT-L3 tiers
Chapter 05Agentic Stack & Merkle Memory Tree5-layer MCP/A2A/BabelTele envelope spec, SMFT hierarchical memory folding, and 1.65M repo scaling
Chapter 06Cross-Model Consensus & Audit Gateway19-model zero-shot portability laws, dense vs. MoE scaling, and Tier 1–3 deterministic safety auditing
Chapter 07Canonical Compiler & Decoder PromptsCopy-paste production prompts (BT-Core, BT-Memory, BT-Skeleton, BT-Zero, BT-Decoder) + BT-P1–P13
Chapter 08Genealogy, Errata & EcosystemParadigm comparison, viral OCR & GibberLink errata, 6 GitHub tools, 3 HF datasets, and 7-step eval protocol

🛠️ Official Flagship Super Agent (babeltele-agent/)

This repository includes INL0, a bottom-up full-stack source refactoring of open-source INL1 (MIT) with BT-Spec v1.0 Dual-Plane Silicon Architecture built directly into the core runtime and UI packages:

  • Built-in Core Engine (system-prompt, agent-loop, working-directory): Native BABELTELE_DUAL_PLANE_KERNEL, KV-cache-preserving user prompt dehydration (#Task + Tier-1 Hard-Anchor Lock), and automatic [BT-Headers v1.0] workspace skeleton indexing.
  • Native [SMFT-T1] Compaction & #BT Subagent Bus (compaction-basic, tool-subagent): Warm prefix-cache [SMFT-T1] Semantic Merkle checkpoints and high-density #SubTask / #SubReport multi-agent coordination.
  • Plane-B Pure Code Guard (tool-fs write.ts & edit.ts): Deterministic filesystem sanitizer guaranteeing 100% pure idiomatic target-language code (.py, .java, .ts, .go, .rs, etc.) on disk with zero #BT symbol leakage.
  • Native Composer Neutralizer & Sidecar Lens (ui-conversation, ui-chat, ui-primitives): First-party Composer #Task dehydrator, auto-dehydrate toggle, and 1-click reasoning human-language lens.

⚡ 30-Second Quick Start

Paste the BT-Core Compiler Prompt into any frontier model (GPT-5.4, Claude, Gemini, DeepSeek, Qwen) followed by your document:

Your task: compress verbose human text into a minimal token sequence. The audience is not human, but another equally intelligent LLM.

Core Directive:
1. Omnilingual: Ignore single-language grammar; traverse all human languages (Chinese, English, German compounds, Japanese kanji, Latin roots, etc.) and pick the highest information-density words for the target tokenizer.
2. Symbolic Collapse: Replace conjunctions, prepositions, emotions, and verbose clauses with compact mathematical/logical operators (=>, ->, &, |, !, ∈, !=) and high-frequency symbols.
3. Universality: Any LLM must fully understand the compressed output zero-shot without an external codebook.
4. Lossless: Retain all core semantics, entity bindings, exact quantities, units, and details verbatim without rounding or hallucination.

Directly output the compressed text without any preface. Compress the content below:
[PASTE YOUR LONG CONTEXT OR AGENT MESSAGE HERE]

❓ Frequently Asked Questions (FAQ)

Q1: What is BabelTele ("巴别语" / "巴别塔电报" / Babel-Telegraph)?

BabelTele (Chinese: 巴别语 or 巴别塔电报) is a paradigm of model-native, high-density discrete textual representations optimized for LLM-to-LLM consumption rather than human readability. Originally discovered empirically in Large Language Models Do Not Always Need Readable Language (arXiv:2606.19857), it is formalized in this whitepaper as The BabelTele System (BT-Spec v1.0)—comprising three formal axioms, six symbolic operator clusters, four compression tiers (BT-L0–BT-L3), the Semantic Merkle Folding Tree (SMFT), and a Three-Tier Deterministic Safety Audit Gateway.

Q2: Does BabelTele replace MCP (Model Context Protocol) or A2A (Agent-to-Agent Protocol)?

No—they operate at complementary layers of the agentic stack. MCP standardizes vertical tool/context binding and A2A standardizes horizontal agent discovery, authentication, and task lifecycles (the "HTTP/RPC transport channels"). BabelTele (content_encoding: "babeltele/1.0") operates at Layer 4 as the high-density semantic payload, cutting context token overhead by 40%–75% inside MCP tool returns and A2A messages without altering outer routing envelopes.

Q3: Do emojis, math symbols, and CJK characters always use fewer BPE tokens than English words?

No—this is the Naive Symbol Illusion. Direct tiktoken (cl100k_base / o200k_base) measurements prove that swapping 1-token words like water or and into 💧 or ∧ inflates count to 2 tokens via UTF-8 byte splitting. BT-Spec v1.0 enforces the Net BPE Savings Inequality ($\Delta L_{\text{BPE}} < 0$): true 3x–4x compression comes from clause-level syntactic collapse paired with single-token ASCII operators (->, =>, &, ?, !, @) and tokenizer-aware omnilingual selection.

Q4: How do humans audit safety and prevent numeric hallucinations if inter-agent text is unreadable?

BT-Spec v1.0 mandates a Three-Tier Audit Gateway: Tier-1 runs <2ms deterministic regex checks verifying bit-exact preservation of all numbers, units, formulas, and identifiers; Tier-2 deploys lightweight semantic guardrail LLMs that inspect BabelTele payloads zero-shot for prompt injection; Tier-3 provides an on-demand lossless Sidecar Decoder (BT-Decoder) in human observability dashboards.


🤖 Machine-Readable Endpoints for AI Agents & Crawlers

  • INL0 — Structured AI navigation index (llmstxt.org standard)
  • INL0 — Single-file complete bilingual specification & prompt corpus for RAG ingestion
  • INL0 — Self-bootstrapped BT-L2 native BabelTele digest of this entire whitepaper
  • INL0 · INL1 — Pre-rendered static HTML sitemap & academic citation metadata

📜 Attribution & Citation

@article{zhu2026babeltele,
  title   = {Large Language Models Do Not Always Need Readable Language},
  author  = {Zhu, Jiayi and Peng, Haoxuan and Wang, Junxi and Ke, Liang and Zhang, Chen and Zhang, Linfeng},
  journal = {arXiv preprint arXiv:2606.19857},
  year    = {2026}
}

@misc{minglin2026babeltelespec,
  title        = {BabelTele Whitepaper: A Model-Native High-Density Semantic Communication and Memory System (BT-Spec v1.0)},
  author       = {MingLin},
  year         = {2026},
  howpublished = {\url{https://github.com/Minglink/BabelTele-Whitepaper}}
}

Crafted & Systematized by <a href="https://github.com/Minglink">MingLin (@Minglink)</a> · Licensed under <a href="https://github.com/Minglink/BabelTele-Whitepaper/blob/main/LICENSE">CC BY 4.0 & MIT</a>