This paper has been superseded.
The content of this paper and Selective Disentanglement (AI-29) have been consolidated into a single paper:
Entanglement Under Fine-Tuning: Scale-Dependent Disentanglement, Crosstalk-Guided Companion Selection, and the Creative Writing Falsification
DOI: 10.5281/zenodo.19430944
Original Executive Summary
This paper tests whether the crosstalk structure of transformer activation space predicts cross-domain interference during fine-tuning. Crosstalk is measured across six domain pairs in Qwen 2.5-Coder-7B: code–creative-writing is nearly block-diagonal (crosstalk 0.0015), while code–legal is heavily coupled (0.30). Fine-tuning Qwen 2.5-Coder-32B under six conditions, the low-crosstalk math companion (B2) beats the natural-language companion (B3) under strong intervention: HumanEval+ 48.2% versus 36.6% (p = 0.016). Complement-subspace regularization (CSR) is the best fine-tuned condition (49.4%), and sequential math-then-code training is catastrophic (10.4%).
An 8-seed strong-intervention replication reveals that B3 (code+NL) drives entanglement intensity to zero in all seeds of Qwen-32B — a complete phase transition. Cross-family replication on CodeLlama-7B and DeepSeek-Coder-6.7B shows this collapse is Qwen-specific: both non-Qwen models maintain or increase EI under the same protocol. Probe validation with three independent sets (including 132 real-dataset probes) confirms the collapse is genuine geometric destruction, not a measurement artifact.
Key References
- McEntire (2026) — Structural Entanglement (AI-26): establishes the geometric phenomenon this paper exploits
- Dettmers et al. (2023) — QLoRA: Efficient Finetuning of Quantized Language Models (NeurIPS)
- McEntire (2026) — The Entanglement Theorem (AI-27): formal proof of the geometric mechanism
- Hu et al. (2022) — LoRA: Low-Rank Adaptation of Large Language Models