01_intro · 完整稿件源码

% 01_intro.tex -- section 1, budget 0.8 page including Fig. 1 (PAPER_PLAN.md float-inclusive budget). Polished 2026-09-02 (hedges and roadmap detail cut).
\section{Introduction}
\label{sec:intro}

\begin{figure*}[t]
  \centering
  \IfFileExists{../figures/fig1.pdf}{\includegraphics[width=\textwidth]{../figures/fig1.pdf}}{\fbox{\parbox[c][5.5cm][c]{0.97\textwidth}{\centering Placeholder for Fig.~1 --- \code{figures/fig1.pdf} is not built yet (source \code{figures/fig1.tex}, pgfplots).}}}
  \caption{One epoch produces most of the dense model's behavioral convergence on \SWEPro (731 tasks); SFT cuts median input by up to 53\,\%, while the capability response depends on the starting model. \textbf{(a)}~Bash-vocabulary Jensen--Shannon distance to the pooled teacher reference (left, log; filled markers) and \passat (right; open markers) over \vTwelve epochs for a post-trained dense model (\QwenDense), a post-trained MoE (\QwenMoE) and a bare MoE base (\QwenMoEBase); dashed line = the natively post-trained \QwenNext. \textbf{(b)}~\passat against median cumulative input tokens per task (log): the dense model moves from 3.27\,M to 1.52\,M at $+2.7$\pp, the bare base from 3.40\,M to 2.44\,M at $+8.9$\pp, and the generation step from 3.27\,M to 4.41\,M at $+5.6$\pp.}
  \label{fig:hero}
\end{figure*}

Open coding agents are increasingly trained on the trajectories of a frontier agent \citep{pan2024training,yang2025swesmith,ahmad2026openswetraces}, and the recipes are compared by resolve rate.
Resolve rate hides the process that produced it \citep{shu2026what}.
At 731 and 500 tasks a paired comparison cannot separate ``the trajectories did not transfer'' from ``they transferred and bought no score'', and two runs of the same weights flip more than a tenth of the tasks.

Imitation SFT is known to copy a model's style before its capability \citep{gudibande2023the}, and agent behaviors can be fingerprinted and made more teacher-like \citep{oderinwale2026agent,yang2026when,yang2026what}; we show where that convergence saturates, what operational habits and input-token footprint it transfers, and why it can be sharply dissociated from paired solve-rate gains on post-trained coding models.
We vary the dose (epochs) and the starting model (bare or post-trained) together, pair every behavioral measurement with a capability measurement on the same arm, add two same-pool anchors (held-out training-pool tasks; the teacher run on the benchmark) and the next Qwen generation as a control, all in one coding harness, and document the curation rule by rule.

The corpus is 8,137 \Teacher/\Harness trajectories (\vTwelve) selected from 32,830 captured runs by an explicit, fully reported rule set (\S\ref{sec:corpus}).
Two open Qwen families are fine-tuned on it from a post-trained dense base, a post-trained mixture-of-experts (MoE) base and a bare MoE base, and twelve arms run in one harness on \SWEPro and \SWEVerified (\S\ref{sec:setup}).
One extractor measures all 46,542 teacher and student trajectories on four rulers, 345 held-out and 359 benchmark tasks anchor teacher and students on the same tasks, and 180 endings and 68 rationalization windows are audited (\S\ref{sec:inherit}--\S\ref{sec:measurement}).

On the dense line on \SWEPro, one epoch produces most of the observed convergence in command vocabulary, tool composition and protocol-error recovery; the second epoch adds a small vocabulary step and the third nothing above the sampling floor (Figure~\ref{fig:hero}a), on the evaluation pools and on the same tasks the teacher ran (Figure~\ref{fig:anchors}).
Supervised fine-tuning (SFT) cuts median cumulative input tokens by 45--53\,\%, from 3.27\,M to 1.52--1.80\,M per task, on every dense checkpoint whatever its score.
Fingerprint proximity changes whether or not \passat changes: the same corpus lifts a bare base by $+8.9$\pp on Pro ($+3.4$\pp on Verified) and moves post-trained bases by $-2.2$ to $+2.8$\pp, below the detectable effect of the paired design, while the next Qwen generation gains $+5.6$ to $+5.8$\pp over the native dense model at $1.35\times$ its input (Figure~\ref{fig:hero}b).
Where paired \passat cannot separate the arms, the behavioral rulers separate them by seven to thirty times their sampling floors (\S\ref{sec:measurement}).

Our contributions are:
\begin{enumerate}
  \item a documented curation pipeline for frontier-agent trajectories with a rule-by-rule impact analysis (\S\ref{sec:corpus});
  \item a dose-response fingerprint analysis: a one-epoch result on the dense line, MoE-line convergence, within-language replication, and two same-pool anchors on which the student's reads and edits overlap the teacher's on tasks it never trained on (\S\ref{sec:inherit});
  \item the dissociation of fingerprint transfer from capability across starting points, with the input-token footprint as a transferred trait (\S\ref{sec:costs});
  \item a paired detectable-effect analysis for \SWEPro \citep{deng2025swebench} and \SWEVerified \citep{jimenez2023swebench,openai2024swebenchverified}, and an audit showing that a success-claim regex tracks completion phrasing rather than outcome (\S\ref{sec:measurement}).
\end{enumerate}