Repository Public Research preview
1% ============================================================================2% UQO Working Paper Template — Economics / Finance3% ============================================================================4\documentclass[12pt,letterpaper]{article}56% --- Encoding & Language ---7\usepackage[utf8]{inputenc}8\usepackage[T1]{fontenc}9\usepackage[english]{babel}1011% --- Page Layout ---12\usepackage[13 letterpaper,14 top=1in,15 bottom=1in,16 left=1in,17 right=1in18]{geometry}1920% --- Typography ---21\usepackage{mathptmx} % Times-like font (text + math)22\usepackage{microtype} % Improved typographic refinements23\usepackage{setspace}24\onehalfspacing % 1.5 line spacing2526% --- Math ---27\usepackage{amsmath,amssymb,amsthm}28\newtheorem{proposition}{Proposition}29\newtheorem{conjecture}{Conjecture}3031% --- Tables ---32\usepackage{booktabs}33\usepackage{tabularx}34\usepackage{multirow}35\usepackage{threeparttable}36\usepackage{siunitx}37\sisetup{38 input-symbols = {()},39 table-format = -1.3,40 parse-numbers = false41}4243% --- Figures ---44\usepackage{graphicx}45\usepackage{url}46\urlstyle{tt}47\usepackage{subcaption}48\usepackage{float}4950% --- Colors & Links ---51\usepackage[dvipsnames]{xcolor}52\usepackage[53 colorlinks = true,54 linkcolor = NavyBlue,55 citecolor = NavyBlue,56 urlcolor = NavyBlue,57 breaklinks = true58]{hyperref}5960% --- Bibliography ---61\usepackage[authoryear,round,semicolon]{natbib}62\bibliographystyle{aer}6364% --- Captions ---65\usepackage[66 font = small,67 labelfont = bf,68 labelsep = period,69 justification = justified,70 singlelinecheck = false71]{caption}7273% --- Headers & Footers ---74\usepackage{fancyhdr}75\pagestyle{fancy}76\fancyhf{}77\renewcommand{\headrulewidth}{0pt}78\fancyfoot[C]{\thepage}7980% --- Section Formatting ---81\usepackage{titlesec}82\titleformat{\section}{\large\bfseries}{\thesection.}{0.5em}{}83\titleformat{\subsection}{\normalsize\bfseries}{\thesubsection.}{0.5em}{}84\titleformat{\subsubsection}{\normalsize\itshape}{\thesubsubsection.}{0.5em}{}8586% --- Appendix Support ---87\usepackage[toc,page]{appendix}8889% --- Misc ---90\usepackage{enumitem}91\setlist{nosep,leftmargin=*}9293% ============================================================================94% METADATA — Edit these fields for each working paper95% ============================================================================96\newcommand{\WPnumber}{AI-0001}97\newcommand{\WPtitle}{The Neural Organ Transplant Experiment}98\newcommand{\WPsubtitle}{Causal Transfer of Cognitive Capabilities Across Local Large Language Models}99\newcommand{\WPdate}{November 2026}100\newcommand{\WPversion}{1.0}101\newcommand{\WPabstract}{%102Can a bounded, causally identified subsystem of one language model be transplanted into another model that103lacks the corresponding capability, without training? We answer in a setting where the question is identifiable:104decoder transformers trained from scratch on an invented operator language, a recipient that never saw one105operator, interchange interventions that localize where the donor executes it, and six transplant families tested106against sham, random, shuffled and unrelated-donor controls and against conventional learning on the same items.107In a preregistered confirmatory run, a donor-in-the-loop activation graft---one residual row at the read position,108mapped through an orthogonal Procrustes transform fitted on task-free data---lifts the recipient from chance to1090.966 on held-out programs (control-adjusted effect $+0.875$, lower bound above the smallest effect of interest in1109 of 9 donor--recipient pairs, control stratum 0.992 with one pair below the 0.95 retention criterion), carries111the rule to unseen operands, needs eight of 128 dimensions, survives width, depth and composition-depth112mismatches, and is anchored by ablation and rescue. The boundary is as informative as the effect: maps fitted by113regression toward the recipient annihilate exactly the capability-specific directions (we prove this for ridge114and CCA), donor-free adapters cache rather than compute, weight grafts damage retained skills, steering vectors115transfer nothing, and a supervised adapter of equal size is perfect. On Qwen2.5-1.5B/3B/7B and Llama-3.2-3B the116same protocol finds task vectors that are causally sufficient within a model for antonyms and capitals, transfers117them between models above null controls in some pairs and directions under a frozen selection rule, and fails for118tasks that no model represents as a vector. The contribution is a causal, preregistered and fully controlled119account of what activation-level transplantation can and cannot move.120}121\newcommand{\WPkeywords}{mechanistic interpretability, activation patching, model stitching, representation alignment, knowledge transfer, causal abstraction, open-weight language models}122\newcommand{\WPjel}{C45, C63, C88, O33}123124% --- Author ---125\newcommand{\WPauthor}{Simon-Pierre Boucher}126\newcommand{\WPaffiliation}{%127 Département des sciences administratives\\128 Université du Québec en Outaouais%129}130\newcommand{\WPemail}{simon-pierre.boucher@uqo.ca}131\newcommand{\WPaddress}{%132 Gatineau -- Pavillon Alexandre-Taché\\133 283, boulevard Alexandre-Taché\\134 Gatineau, Québec, Canada J9A 1L8%135}136137% ============================================================================138% DOCUMENT139% ============================================================================140\begin{document}141142% --- Title Page ---143\input{sections/titlepage}144145% --- Main Body ---146\input{sections/introduction}147\input{sections/related_work}148\input{sections/operational_theory}149\input{sections/experimental_design}150\input{sections/results}151\input{sections/mechanistic_analysis}152\input{sections/robustness}153\input{sections/discussion}154\input{sections/limitations}155\input{sections/conclusion}156157% --- References ---158\newpage159\bibliography{references}160161% --- Appendix ---162\input{appendix/appendix}163164\end{document}165