Our AGI Research Division focuses on scalable reasoning, value alignment, multimodal cognition, and open benchmarks. We publish transparently, collaborate globally, and release all datasets and code under permissive licenses.
Our AGI program is structured around four foundational tracks, each addressing critical bottlenecks in path-to-general intelligence systems.
Architectures that enable slow, deliberate, step-wise reasoning with self-correction, uncertainty quantification, and causal inference.
ActiveRobust preference learning, constitutional AI frameworks, and red-teaming protocols to ensure corrigibility and human-compatible objectives.
ActiveUnified representations across text, vision, audio, and embodied simulation for grounded, transferable understanding.
ActiveReproducible, contamination-resistant evaluation suites measuring reasoning, creativity, and cross-domain transfer without leaderboard gaming.
In ProgressPeer-reviewed and pre-print research from the Aevum AGI Division. All code and supplementary materials are open-source.
We introduce a dynamic pruning mechanism that selectively collapses redundant reasoning steps while preserving logical validity, reducing inference latency by 42% without accuracy degradation on GSM8K and MMLU-Pro.
A novel reward modeling framework that integrates culturally-diverse preference datasets with constitutional constraints, demonstrating improved robustness against adversarial prompt injection and value misalignment.
We show that injecting physics-informed simulation priors into multimodal transformers significantly improves spatial reasoning and causal prediction in complex visual-language tasks.
Freely available for academic and commercial research. All releases include documentation, citation guidelines, and usage licenses.
| Dataset / Benchmark | Domain | Size / Scope | License | Access |
|---|---|---|---|---|
| Aevum-Reason-1M | ReasoningLogic | 1.2M samples | CC-BY-4.0 | Download |
| CrossCult-Align-50K | AlignmentPreferences | 50K pairs, 18 regions | CC-BY-NC-4.0 | Download |
| MultiSim-Embodied | MultimodalPhysics | 850K trajectory clips | Apache-2.0 | Download |
| Aevum-AGI-Bench v2.1 | EvaluationRolling | 12K dynamic tasks | MIT | Leaderboard |
Interdisciplinary scholars, ML engineers, and cognitive scientists driving our AGI initiative forward.
Receive monthly research digests, dataset releases, and early access to pre-prints. We share our methodology, failures, and breakthroughs transparently.
For collaboration inquiries, contact: agi-research@aevum.edu