A complete reference for all 38 cognitive modules: algorithms, inputs, outputs, and the neuroscience research behind each one.
Modules organized in concentric layers, from core cognition to AGI extensions.
20 modules implementing core cognitive capabilities
6 modules providing system-level services
5 modules for evaluation, perception, and integration
7 modules for advanced AGI capabilities
Evaluation
627 tests across 5 categories. All passing.
| Category | Score | Distribution | Tests | Status |
|---|
Counterfactual reasoning, structural causal models, intervention calculus, and causal DAG discovery. 124 tests covering Pearl's causal hierarchy.
Neuro-symbolic binding, hierarchical concept composition, ARC-style pattern generalization. 144 tests on compositional generalization.
Elastic weight consolidation, hippocampal replay, meta-learning. 89 tests verifying no catastrophic forgetting across task sequences.
Uncertainty quantification, out-of-distribution detection, adversarial defense, probability calibration. 192 tests under distribution shift.
Semantic parsing, pragmatic inference, discourse modeling, dialogue management. 78 tests on comprehension and generation quality.
Weighted average across all categories. 627 total tests, all passing. Comparable to or exceeding established cognitive architectures on structured reasoning tasks.
How ELO-AGI compares to established cognitive architectures and LLMs.
| Capability | ELO-AGI | ACT-R | SOAR | GPT-4 |
|---|---|---|---|---|
| Total Modules | 38 | ~12 | ~8 | Monolithic |
| Causal Reasoning | 0.850 | 0.420 | 0.380 | 0.610 |
| Compositional Gen. | 0.740 | 0.310 | 0.290 | 0.520 |
| Continual Learning | 0.720 | 0.180 | 0.150 | 0.000 |
| Robustness | 0.710 | 0.350 | 0.320 | 0.440 |
| Language | 0.690 | 0.250 | 0.220 | 0.890 |
| Overall | 0.742 | 0.302 | 0.272 | 0.492 |
Key papers and theories informing ELO-AGI's architecture.
Foundation for the attention-gated broadcasting mechanism. Modules compete for access to the global workspace; winners broadcast their content to all other modules simultaneously.
Core inference mechanism. The system minimizes free energy through hierarchical prediction and prediction error propagation, implementing variational Bayesian inference.
System 1 (fast, intuitive) and System 2 (slow, analytical) reasoning pathways. Override detection triggers deliberate reasoning when intuitive confidence is low.
Extends the Free Energy Principle to action selection. The system selects actions that are expected to minimize future prediction error, unifying perception, learning, and decision-making.
Emotional signals from body states influence decision-making. The Emotion module generates somatic markers that bias value computation in the Decision Making module.
Survey of agentic AI capabilities informing the multi-agent coordination and self-improvement modules.
Compositional generalization benchmark informing the Compositional Abstraction module design and evaluation methodology.