NE-OS Research
Publications
Independent AI research from NE-OS SpA. Focus areas: novel LLM architectures, mechanistic interpretability, and latent space geometry.
4 preprints·June 2026·founder@ne-os.com
AION: Adaptive Inference with Orthogonal Networks
Rodrigo Campos Vargas — NE-OS SpA, Chile · June 2026
ArchitectureEfficient InferenceMemory
→99.6% / 0.0% factual/generative gate separation
→Hard Bypass: zero transformer FLOPs for factual queries
→Two-level efficiency: Hard Bypass + per-layer Early Exit
DOI: 10.5281/zenodo.20757442
RCLA: Resonance-Coded Language Architecture
Rodrigo Campos Vargas — NE-OS SpA, Chile · June 2026
ArchitectureSparse RoutingEfficiency
→13.4% lower loss than transformer baseline
→1.61x faster training
→Path to CPU inference via binary popcount similarity
DOI: 10.5281/zenodo.20757444
Universal Latent Space Alignment Across LLM Families via Linear Projection
Rodrigo Campos Vargas — NE-OS SpA, Chile · June 2026
Representation LearningInterpretabilityAlignment
→cos > 0.99 across Mistral, Falcon, Gemma → Qwen
→4MB linear projection sufficient for alignment
→Empirical confirmation of Platonic Representation Hypothesis
DOI: 10.5281/zenodo.20757527
Three Phases of Internal Processing in Transformer LLMs
Rodrigo Campos Vargas — NE-OS SpA, Chile · June 2026
Mechanistic InterpretabilityLLM InternalsAnalysis
→3 universal processing phases in all transformer LLMs
→False vs. real token emergence distinction
→Training corpus signature detectable in Phase 1
DOI: 10.5281/zenodo.20757528