Na minha lista:
| Autor principal: | |
|---|---|
| Formato: | Recurso digital |
| Idioma: | |
| Publicado em: |
Zenodo
2026
|
| Assuntos: | |
| Acesso em linha: | https://doi.org/10.5281/zenodo.20043990 |
| Tags: |
Adicionar Tag
Sem tags, seja o primeiro a adicionar uma tag!
|
Sumário:
- <p class="MsoBodyText">Abstract</p> <p class="MsoBodyText">The pursuit of Artificial General Intelligence (AGI) has been dominated by inductive, probabilistic approaches that teach models <em>what</em> to think through massive parameter scaling and pattern memorization. We present the PACAD (Profile, Axiomatic Canonical, and Domain) three-tier knowledge verification architecture as a foundational alternative that teaches models <em>how</em> to think through deductive structural reasoning. Moving beyond the theoretical distinctions established in prior work, we formalize PACAD as an operational blueprint for systematically escalating raw information (Profiles) through rigorous validation into Tested, Verified Canonicals (TVCs), and ultimately into axiomatized, domain-independent Canonicals expressed through the Fr(N,μ,D) relational grammar. This hierarchical progression provides a formal criterion for valid reasoning steps that addresses a critical gap in existing Process Reward Models. We demonstrate how this architecture enables cross-domain structural transfer without task-specific fine-tuning, and we present evidence from recent industry developments—including the anomalous performance trajectory of Anthropic’s Claude Mythos (April 2026)—as empirical support for the structural reasoning hypothesis. We propose Oracle, a full PACAD realization, as the successor to partial implementations, and we advance three falsifiable predictions: (1) a PACAD-grounded model will outperform comparably-sized LLMs on ARC-AGI tasks requiring cross-domain structural transfer without task-specific fine-tuning; (2) the TVC standard provides a formal, automatable criterion for reasoning step validity absent from current Process Reward Models; (3) canonical composition scales sub-linearly in parameters for structural reasoning tasks, whereas flat-vector architectures exhibit diminishing returns. The PACAD architecture represents not merely an epistemological framework but a critical risk mitigation strategy for multi-billion dollar AGI development, ensuring verifiable, explainable, and robust general intelligence.</p>