Published March 12, 2026 | Version v2
Journal article Open

Entanglement-Optimal Fine-Tuning: Leveraging Structural Concept Coupling for Parameter-Efficient Adaptation

Authors/Creators

  • 1. Cage & Mirror Press

Description

Structural entanglement predicts fine-tuning interference proportional to crosstalk between domains. 8-seed strong-intervention replication shows B3 (code+NL) drives entanglement intensity to zero in all Qwen-32B seeds — a sharp phase transition at steps 2000-3500. Cross-family replication on CodeLlama-7B and DeepSeek-Coder-6.7B confirms the collapse is Qwen-specific. Probe validation with three independent sets (including 132 real-dataset probes) confirms the collapse is genuine geometric destruction. CSR (complement-subspace regularization) preserves structure with 95% CI spanning zero.

Notes

Paper 44 in the Structural Compression Theory research program by Jeremy McEntire. V2: Added 8-seed strong-intervention replication, cross-family replication (CodeLlama, DeepSeek), and probe validation. Renamed from Universal to Structural entanglement.

Files

entanglement_optimal_coding.pdf

Files (355.3 kB)

Name Size Download all
md5:2495a3fb303593c643a2a00573c699c7
355.3 kB Preview Download