Single-Dim Functional Control and Pathway-Specific Inhibition in a Passthrough-Merged Qwen2.5 Hybrid
Year: 2026
Type: preprint
Abstract: © Copyright 2026 Chen, Ho Yiing (ORCID: 0009-0006-6816-9891), Independent researcher, charenix.com. All rights reserved by the author. This work is licensed under the Creative Commons Attribution 4.0. Restoring exactly one hidden-state coordinate (d=758, last four blocks of a passthrough-merged qwen2.5 hybrid) flips a Chinese narrative response into a bilingual codeblock-with-emoji response on a specific technical-explanation input, while leaving four other inputs visually unchanged. Four findings: Single-d controller demo: Restoring d=758 alone in blocks 24-27 of v7b (a 24-block Instruct + 4-block Coder passthrough merge) triggers a reproducible style switch (Chinese narrative to bilingual codeblock + crab emoji) on the race-condition prompt. Other four prompts unchanged. 11-dim functional taxonomy: Sweeping one-at-a-time restoration across the 11 Mercury anchor dimensions sorts them into 4 categories: clean triggers (279, 382, 476, 758), trigger+destabilizers (11, 334, 510), pure destabilizers (12, 25, 481), and one hidden inhibitor (715). Pathway-specific inhibition: Pairing d=715 with each trigger reveals a binary pathway split: {279, 476} fully suppressed by 715, {382, 758} immune. First reported structural relationship inside the 11 anchors. Subspace consistency principle: Layer-norm-only ablation (9 scalar overrides) collapses generation on 2 of 5 prompts; full subspace restoration (50+ tensor slices per anchor) preserves coherence. Steering edits need full read-pathway integration, not scalar gain edits. Single-coordinate control reached by ablation alone, no sparse autoencoder training, no gradient-based feature attribution. The controller is one integer index in the model's native basis. Companion records: Mercury method (10.5281/zenodo.20313154), discovery (10.5281/zenodo.20313748), viewer dataset (10.5281/zenodo.20313150), scaling (10.5281/zenodo.20325676), cross-model dim preservation (10.5281/zenodo.20346626), SLERP/TIES case study (10.5281/zenodo.20346628).
Author: Chen, Ho Yiing
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7162113442
Cites:
Cited by:
Related to: