现稿摘要与章节入口

% main.tex -- COLING 2027 long paper (Macau, 9-14 May 2027; ARR October 2026 cycle, deadline 2026-10-12 AoE), anonymous review build. ACL template = ARR template.
% Build: `make` in this directory (pdflatex -> bibtex -> pdflatex x2), see README.md.
% Style: acl-style-files @ d5adc823 (2026-06-29); acl.sty / acl_natbib.bst copied next to this file.
\documentclass[11pt]{article}

% [review]  : anonymous author block, line numbers, page numbers (ARR submission)
% [preprint]: non-anonymous with page numbers (arXiv)
% no option : camera-ready
\usepackage[review]{acl}

% Standard template packages (acl.sty itself loads geometry, xcolor, caption, natbib, hyperref)
\usepackage{times}
\usepackage{latexsym}
\usepackage[T1]{fontenc}
\usepackage[utf8]{inputenc}
\usepackage{microtype}
\usepackage{inconsolata}
\usepackage{graphicx}

% Arm / model / corpus names, \todo, booktabs, pgfplots, ...
\input{macros}

% Float placement: the body is float-dense (six figures, one table* in eight pages), so let two
% double-column floats share a page top and let a page be mostly floats; otherwise LaTeX defers
% figures past the section that discusses them.
\setcounter{topnumber}{3}
\setcounter{dbltopnumber}{2}
\renewcommand{\topfraction}{0.95}
\renewcommand{\dbltopfraction}{0.95}
\renewcommand{\textfraction}{0.05}
\renewcommand{\floatpagefraction}{0.8}
\renewcommand{\dblfloatpagefraction}{0.8}

% Working title (PAPER_PLAN.md, reviewer's 10-word proposal)
\title{What Agentic SFT Buys: Rapid Behavioral Convergence, Base-Conditional Capability Gains}

% Replaced by "Anonymous ACL submission" while the [review] option is active.
% Fill in for the preprint / camera-ready build.
\author{Anonymous Author(s) \\
  Anonymous Affiliation \\
  \texttt{anonymous@example.org}}

\begin{document}
\maketitle

\begin{abstract}
% ---- PAPER_PLAN.md section 0 (target ~200 words) ----
% - What we achieve: a curated frontier-agent corpus and a behavioural-fingerprint analysis showing what
%   agentic SFT transfers (habits and token footprint within one epoch) and what it does not
%   (capability on already post-trained bases).
% - Why it matters / is hard: trajectory SFT is the dominant recipe for open coding agents, but scores alone
%   cannot tell "transfer failed" from "transfer worked but bought nothing" at 731/500 tasks.
% - How: 12 arms (dense 27B x1/2/3 epochs and a by-product variant; MoE from bare and post-trained bases;
%   vendor next generation) in one harness; one extractor over 46,542 trajectories; a same-task anchor;
%   audited endings; a documented curation pipeline.
% - Evidence (reviewer's wording): one epoch produced most of the observed dense-model behavioral convergence;
%   SFT cut median input by up to 53 %, improved one bare base by +8.9 pp on Pro (+3.4 pp, inconclusive,
%   on Verified), and yielded -2.2 to +2.8 pp against post-trained baselines, with no exact baseline
%   comparison resolved; a generation jump gave +5.6-5.8 pp at 2.9x cost; 0/110 audited failure endings
%   admit failure.
% - Most remarkable: one epoch produces most of the habits and halves the input tokens; the capability
%   response depends on the starting model.
Open coding agents are increasingly trained on the trajectories of a frontier agent, and the recipes are compared by resolve rate, which at 731 and 500 tasks cannot separate ``the trajectories did not transfer'' from ``they transferred and bought no score''.
We curate 8,137 \Teacher/\Harness trajectories from 32,830 captured runs with an explicit rule set, fine-tune two open Qwen families on them for one to three epochs from bare and post-trained bases, and evaluate twelve arms in one harness on \SWEPro and \SWEVerified; one extractor fingerprints all 46,542 teacher and student trajectories, and two same-pool anchors put teacher and students on the same tasks.
On \SWEPro, one epoch produces most of the dense model's convergence in command vocabulary, tool composition and protocol-error recovery, and on the same tasks cuts the distance to the teacher's own trajectories to about one tenth of the native model's; supervised fine-tuning (SFT) cuts median cumulative input tokens by 45--53\,\% on every dense checkpoint, whatever its score.
Fingerprint proximity changes whether or not \passat changes: the corpus lifts a bare base by $+8.9$\pp on Pro ($+3.4$\pp on Verified), every comparison against a post-trained base ($-2.2$ to $+2.8$\pp) stays below the paired design's detectable effect, and the next Qwen generation gains $+5.6$ to $+5.8$\pp over the native dense model at $1.35\times$ its input.
Where paired \passat at 731 and 500 tasks cannot separate the arms, the behavioral rulers separate them by seven to thirty times their sampling floors.
\end{abstract}

\input{sections/01_intro}
\input{sections/02_related}
\input{sections/03_corpus}
\input{sections/04_setup}
\input{sections/05_inherit}
\input{sections/06_costs}
\input{sections/07_measurement}
\input{sections/08_conclusion}

% Mandatory Limitations section (not counted toward the page limit) and the ethics statement
\input{sections/09_limitations}
\input{sections/10_ethics}

% lit/refs.bib: arXiv-verified entries (lit/make_bib.py); lit/refs_manual.bib: vendor/blog references marked [VERIFY].
\bibliography{references}

\appendix
% Appendix floats are numbered A1, A2, ... so that body figures keep their citation order.
\setcounter{figure}{0}\renewcommand{\thefigure}{A\arabic{figure}}
\setcounter{table}{0}\renewcommand{\thetable}{A\arabic{table}}
\input{appendix/A_evaluation_noise}
\input{appendix/B_fingerprint_robustness}
\input{appendix/C_extras}
\input{appendix/D_reproducibility}
\input{appendix/E_anchors}

\end{document}