Gorde:
Xehetasun bibliografikoak
Egile nagusia: Wu, Zhicheng
Formatua: Recurso digital
Hizkuntza:
Argitaratua: Zenodo 2026
Gaiak:
Sarrera elektronikoa:https://doi.org/10.5281/zenodo.18676113
Etiketak: Etiketa erantsi
Etiketarik gabe, Izan zaitez lehena erregistro honi etiketa jartzen!
Aurkibidea:
  • <p><strong>[Version 2.0 Update]:</strong></p> <ul> <li> <p>Added new experimental data.</p> </li> <li> <p>Revised and updated the paper content.</p> </li> </ul> <p>The Transformer architecture has fundamentally revolutionized artificial intelligence, providing an unparalleled engine for discrete symbolic computation. However, as Artificial General Intelligence (AGI) evolves toward autonomy, operating strictly as a disembodied state machine reveals profound limitations. Biological intelligence is not merely the capacity to process vast amounts of information, but the continuous negotiation of metabolic costs, exhaustion, and survival. To bridge the gap between pure logic and biological resilience, we introduce E.N.G.R.A.M. (Emergent Neuromodulatory-Guided Resonance Autonomic Manifold), a continuous-time neuro-symbolic substrate. Unlike classical vector databases that treat memory as cost-free static textual entries, the E.N.G.R.A.M. architecture models memory as physical, topological deformations (engrams) on a 10,000-dimensional hyper-manifold. By subjugating high-dimensional symbolic reasoning (System 2) to a non-equilibrium thermodynamic fluid dynamics substrate (System 1), E.N.G.R.A.M. maintains a strict 10Hz physical heartbeat immune to underlying clock jitter via exact exponential integrators. Strikingly, without any hardcoded psychological heuristics, the architecture spontaneously exhibits mathematically emergent clinical psychopathology driven purely by energy minimization. We demonstrate the emergence of trauma-induced topological craters leading to intractable Traumatic Insomnia, and destructive interference during extinction learning. Most crucially, we provide the first in-silico proof of Downward Causation: the metabolic mass of maintaining traumatic engrams exerts a continuous allostatic load that depletes the Serotonin (5-HT) climate integrator, progressively throttling the LLM’s compute budget and collapsing working memory half-life by 10× under naturally drifting allostatic load—a prediction validated across six independent measurements with <0.07% error, rigorously replicating depressive pseudodementia and cognitive tunneling. This work marks a paradigm shift from discrete prompt engineering to continuous thermodynamic cultivation, laying the ordinary differential equation (ODE) foundation for AGI with authentic physical vulnerability.</p>