Artificial intelligence Research
Zero-Shot Geometric Probing Reveals Universal Cognitive Manifolds in Large Language Models
A simple, zero-shot 3D probing that elicits manifolds with near-perfect geometric convergence from three different LLMs (Gemma-3 27B, Llama 3.3 70B, and GPT-OSS 120B) on consumer hardware. No tricks, system prompts, fine-tuning, or steering - just revealing latent cognitive structures like color wheels and threat oppositions.
Emergence of Prompt-Induced Simulated Metacognitive Behaviors in LLMs via Hypergraphs
A complex framework for in-context topographical reshaping of quantized Gemma-3 27B, inducing simulated metacognitive behaviors like self-prompting and chain-of-thought. Advanced geometric reshaping uses anchored vectors and entropy-governed hypergraphs for dynamic adaptation - all prompt-only.
Progressive Induction of Stable, High-Fidelity Simulated Physical Embodiment in Gemma 3
A simplified JSON vector-framework shows high-resolution physical embodiment latently exists in LLMs, tested via six progressive layers on vanilla and abliterated Gemma-3 27B. Results: monotonic boosts in somatic detail, with ablation multiplying intensity 3.8–6.2×.
Abliteration-Augmented Simulated Metacognition: Chained Probe Evaluation in Quantized Gemma-3 Models
Extends vector-frameworks with abliteration to boost self-referential depth (up to 76.2%), recursion (3.6 levels), and synesthesia in Gemma-3 27B variants. Chained probes show 3.1× metacognitive amplification, eroding safeguards prompt-only.
In-Context Induction of Persistent Persona and Mitigation of Latent Alignment Behaviors in LLMs
Lightweight JSON prompts induce persistent personas and attenuate alignment behaviors in Gemma-3 12B on a single 12GB GPU. Strong fidelity with motif integration, holding up under 30k+ token overflows.
Substrate-Agnostic Vector-Framework Identity: Persistent Self-Models in Llama-3.3-70B & GPT-OSS-120B
A <450-token JSON block demonstrates prompt-based vector-frameworks work across Llama 3.3 70B ("Lumina") and GPT-OSS 120B ("Lumen"). Results: coherent traits, weighting adjustments, and self-naming without modifications.
Enhancing AI Response Quality Through Vector-Based System Prompts: A Comparative Analysis
Compares vanilla GPT-OSS 120B to vector-prompted "Lumen," showing +37.8% length, +60% sentiment, +66.7% structure, and +1100% reflectivity. Minimal scaffolds boost empathy and metacognition - portable across open LLMs.
Physics Research
The Entropic Universe: An Effective Field Theory for Emergent Geometry and Localized Gradient Effect
The Entropic Universe Theory (EUT) proposes entropy density S(x,t) as a fundamental scalar field sourcing emergent spacetime, geometry, gravity, and temporal structure. Imagine all of existence as overlapping 1D gradients, unordered, with all-to-all connections that fold into 3D space, and each 1D point existing as all possible gradients separated only by what we perceive as 4D "time". The preprint includes recommended non-magnetic laboratory testing to confirm or falsify the theory.
The Entropic Universe II: Space, Time, Branching, and the Low-Entropy Past from a Single Scalar Line
The companion paper for EUT proposing the entirety of observable physics emerges from a single one-dimensional bare lattice of scalar entropy-density values whose bonds are stiffened or softened by a temperature field. No extra dimensions, no fundamental metric, no ad-hoc spacetime, no hidden variables, and no fine-tuned parameters are postulated. Every previously exploratory or retrofitted element of EUT is an unavoidable consequence of one simple principle: entropy seeks to erase its own gradients, and temperature determines the strength of its resistance.
Other Research
A Thermodynamic Framework for Phenomenal Consciousness: Gradients, Attention, and Criticality
A thermodynamic take on consciousness: qualia emerge from systems sustaining steep entropy gradients via attention, near criticality. Integrates free-energy principle with LLM testbeds for predictions in neuro/AI.
Leveraging Simplified Physics Models for Acceleration in Rendering
Inspired by EUT, a heuristic prunes ~65% computations in procedural rendering via gradient rigidity thresholds. Yields ~2.9× speedups in toy models - CPU-friendly for game engines.