| # AETHER: A Self-Evolving Neuro-Symbolic Architecture for AGI |
|
|
| **v0.2.0 — Autonomous Mode** |
|
|
| AETHER (Adaptive Evolving Towards Higher-order Reasoning) is a unified self-evolving neuro-symbolic architecture that integrates symbolic and sub-symbolic computation within a dynamically self-modifying framework. This version runs **fully autonomously** with zero human-in-the-loop oversight — all safety, validation, and rollback decisions are handled by automated systems. |
|
|
| --- |
|
|
| ## Architecture Integration |
|
|
| AETHER synthesizes cutting-edge research from: |
|
|
| | Component | Source | Key Contribution | |
| |-----------|--------|-----------------| |
| | **Evolutionary Core** | AlphaEvolve (DeepMind, 2025) | MAP-Elites + island model + LLM code diffs for algorithm discovery | |
| | **Hierarchical Reasoning** | HiMAC (2026) | Macro-Policy / Micro-Policy co-evolution with iterative optimization | |
| | **Group Self-Evolution** | GEA (2026) | Performance-Novelty selection with experience sharing | |
| | **Tool Evolution** | Yunjue Agent (2026) | Manager/Executor/Developer/Integrator role decomposition + tool absorption | |
| | **AI Research Agent** | ASI-Evolve (2026) | 4-stage Researcher→Engineer→Analyzer→Database loop | |
| | **Co-Evolution** | CoMAS (2025) | Decentralized multi-agent co-evolution via interaction rewards | |
| | **Cognitive Architecture** | CoALA (2023) | Working/Episodic/Semantic/Procedural memory taxonomy | |
| | **Agentic Neural Networks** | ANN (2025) | Textual backpropagation across multi-agent layers | |
| | **Leader Training** | MLPO (2025) | Train single leader, peers untrained - efficient multi-agent RL | |
| | **Task-Driven Agents** | BabyAGI | Task creation / prioritization / execution loop | |
|
|
| --- |
|
|
| ## System Components (v0.2.0) |
|
|
| ``` |
| AETHER |
| ├── Core (AetherCore) |
| │ ├── Neuro-Symbolic Fusion Gate (learned attention weights) |
| │ ├── Recursive Evolution Loop (generate→evaluate→select→mutate→validate→integrate) |
| │ └── AutoOversight System (risk scoring + regression suite + auto-rollback) |
| ├── Memory (CoALA-inspired) |
| │ ├── Working Memory (attention-based retrieval) |
| │ ├── Episodic Memory (experience buffer) |
| │ ├── Semantic Memory (world knowledge via KG) |
| │ └── Procedural Memory (learned tools/skills) |
| ├── Knowledge (PyG-style) |
| │ ├── RGCN Encoder (relational graph convolution) |
| │ ├── ComplEx Scorer (link prediction) |
| │ └── Symbolic Rule Engine (forward chaining + multi-hop BFS) |
| ├── Agents |
| │ ├── Hierarchical Agent (HiMAC: Macro + Micro policy) |
| │ ├── Agent Orchestrator (MLPO leader + dynamic routing) |
| │ ├── BabyAGI Loop (task-driven autonomy) |
| │ └── Textual Backpropagation (Agentic NN update) |
| └── Evolution |
| ├── MAP-Elites Archive (quality-diversity) |
| ├── Performance-Novelty Selection (GEA) |
| ├── Constrained Mutation (AlphaEvolve) |
| └── AutoOversight Gate (replaces human-in-the-loop) |
| ``` |
|
|
| --- |
|
|
| ## What's New in v0.2.0 (Autonomous) |
|
|
| - **AutoOversight System** — Replaces all human-in-the-loop safety gates with automated regression suites, risk scoring, and auto-rollback. The system evaluates its own candidates via synthetic benchmarks before any integration. |
| - **Self-Sustained Evolution** — The evolution loop runs end-to-end without external API calls or human checkpoints. Fitness is evaluated against internal reasoning, memory, and knowledge graph benchmarks. |
| - **Temporal Memory with Attention** — Recency-weighted retrieval for long-horizon context across evolution generations. |
| - **Four-Agent Orchestration** — Researcher, Engineer, Analyzer, Integrator roles with learned routing weights. |
| - **Fully Runnable Standalone** — Single-file executable with only `torch`, `numpy`, and `networkx` dependencies. |
|
|
| --- |
|
|
| ## Quick Start (Autonomous) |
|
|
| ```bash |
| pip install torch numpy networkx |
| python aether_autonomous.py |
| ``` |
|
|
| ```python |
| from aether_autonomous import AetherCore, AetherConfig |
| |
| # Initialize autonomous AETHER |
| config = AetherConfig( |
| population_size=6, |
| generations=5, |
| mutation_rate=0.12, |
| macro_policy_dim=128, |
| micro_policy_dim=64, |
| num_agents=4, |
| kg_embedding_dim=64, |
| kg_num_relations=10, |
| ) |
| |
| aether = AetherCore(config) |
| |
| # Seed knowledge |
| aether.knowledge.add_fact("Intelligence", "requires", "Reasoning") |
| aether.knowledge.add_fact("Reasoning", "requires", "Memory") |
| |
| # Neuro-symbolic query |
| result = aether.forward("Intelligence requires") |
| print(f"Symbolic weight: {result['symbolic_weight']:.3f}") |
| print(f"Neural weight: {result['neural_weight']:.3f}") |
| |
| # Fully autonomous evolution (no human oversight) |
| evolution_result = aether.evolve(num_generations=5) |
| print(f"Best fitness: {evolution_result['best_fitness']:.4f}") |
| print(f"Archive coverage: {evolution_result['archive_stats']['coverage']:.2%}") |
| ``` |
|
|
| --- |
|
|
| ## Original Modular System (v0.1.0) |
|
|
| The original modular implementation remains available in the `aether/` directory: |
|
|
| ```python |
| from aether.core import AetherCore, AetherConfig |
| |
| config = AetherConfig( |
| population_size=8, |
| mutation_rate=0.15, |
| num_agents=4, |
| enable_self_modification=True, |
| ) |
| aether = AetherCore(config, model_name="Qwen/Qwen2.5-0.5B-Instruct") |
| ``` |
|
|
| --- |
|
|
| ## Design Principles |
|
|
| 1. **Neuro-Symbolic Fluidity**: Dynamic translation between symbolic and sub-symbolic representations |
| 2. **Architectural Evolvability**: Structural components are subject to learning and refinement |
| 3. **Parallel Agent Intelligence**: Intelligence emerges through coordinated multi-agent interaction |
| 4. **Constrained Self-Modification**: All self-changes are sandboxed and validated by automated systems |
| 5. **Automated Oversight**: Risk scoring, regression suites, and auto-rollback replace human gates |
|
|
| --- |
|
|
| ## Citation |
|
|
| ```bibtex |
| @article{aether2026, |
| title={AETHER: A Self-Evolving Neuro-Symbolic Architecture for Artificial General Intelligence}, |
| author={Anonymous}, |
| year={2026} |
| } |
| ``` |
|
|
| ## License |
|
|
| MIT License - Open for research and development toward responsible AGI. |
|
|
