An experimental, CPU-runnable research scaffold for bounded closed-loop learning and validation-gated recursive self-improvement.
DeepNeural-AutoExploration is a structured sandbox designed to inspect and evaluate bounded neural architecture mutations:
- Multiprocessing Candidate Evaluation: Evaluates mutated proposals concurrently in process-isolated workers.
- Robust Fault Tolerance: Intercepts training, mathematical instability (NaNs), and dimensional mismatch failures to gracefully roll back failed candidates.
- Evidence-Guided Deep Search: Automatically schedules search depth and mutations using representation/gradient probes and rollback histories.
- Meta-Meta-Meta Controller: Self-diagnoses the search process and proposes optimized next-run configurations.
Selected same-shape ARC-AGI-1 results:
| Benchmark | Mode/seed | Baseline | Evolved | Delta |
|---|---|---|---|---|
| ARC cell accuracy | full/42 | 0.668 |
1.000 |
+0.332 |
| ARC exact-grid accuracy | full/42 | 0.000 |
1.000 |
+1.000 |
| ARC exact-grid accuracy | quick/42 | 0.000 |
0.400 |
+0.400 |
- Supported: Bounded model candidate mutation, training, probing, evaluation, and rollback on validation tasks. The repository allows the user to inspect loops that generate bounded operator programs.
- Unsupported: This repository is not AGI, not human-level intelligence, and not a technological singularity system. It does not prove open-ended autonomous recursive self-improvement. It is a research scaffold for testing code-level mutations and operator programs under verification rules.
git clone https://github.com/sunghunkwag/DeepNeural-AutoExploration.git
cd DeepNeural-AutoExploration
pip install -r requirements.txt pytestRun core regressions:
pytest -qRun exploration loops:
# Autonomous neural exploration loop
python benchmarks/autonomous_neural_exploration_loop.py --mode smoke --seed 42
# Recursive self-improvement (RSI)
python benchmarks/recursive_self_improvement_benchmark.py --mode smoke --seed 42
# Long-horizon neural exploration
python benchmarks/long_horizon_autonomous_exploration.py --mode full --seed 42 --runs 10neural_search/: Bounded architecture mutations, weight inheritance, task families, and activation/gradient probes.failure_residue/: Structured failure residue extraction and leak refusal checks.meta_rsi/: Orchestrates experiment plans, registry updates, and evaluator repairs.world_model_v2/: Object-centric state representations, causal graphing, and rollouts.benchmarks/&tests/: Bounded loop execution runners and regression tests.
For detailed analysis, reports, and plans, refer to:
- External RSI Evidence Report
- Research-Goal Generation Integration
- Next Architecture Upgrade Plan
- Deep Neural Autonomous Exploration Upgrade
- Long-Horizon Autonomous Exploration Report
- Long-Horizon Seed-42 Failure Analysis
- Long-Horizon Repair & Memory Ablation Report
- ARC Seed-44 Failure Residue Report
- Interaction Operating Scaffold
- Interaction Residue Layer
Apache License 2.0.