Help AI Understand Your Whole Paper: Build a Full-Paper Index Before Asking Questions
How can you tell whether AI has understood the whole paper? Ask a question that requires connecting sections, such as: “How do the assumptions in Section 3 support the conclusions in Section 5?” If it merely repeats one section, it has not yet connected the argument across the paper.
In EasyLaTeX, open the relevant project when you want to analyze the LaTeX paper you are writing. In the project's AI conversation, explicitly ask it to read the project files and map the structure without changing the text. Being able to read files does not mean it has already read every section in full on every turn. Ask which files it actually read and what remains unread; for a long paper, check that it did not stop after the opening portion. For an external PDF, use AI on the literature reading page. Without LaTeX source, locate evidence by page number, section heading, and equation number rather than asking for source labels that do not exist.
Step 1: Ask for a section-and-label inventory from the source
First, ask AI to map the paper's skeleton: hierarchy, headings, the original \label values in the source, and a one-sentence summary of each section. This inventory gives later conversations a shared coordinate system. Whether you want to discuss an edit or reference a passage, you can name the entry. It is especially useful when a long paper has several subsections with the same heading, such as “Experimental Setup.”
An inventory might look like this. Labels must be copied exactly from the source, never invented:
1 Introduction sec:intro 提出问题与本文贡献
2 Method sec:method 定义步长 \lrate 与目标 \loss
2.1 Update rule sec:update 给出更新式 (eq:update)
3 Convergence sec:conv 在假设 A 下给出步数上界
4 Experiments sec:exp 用两种步长 schedule 检验第 3 节结论
This assumes the source uses \section and \label. If there are no labels, ask for headings in source order instead. A common pitfall is that AI helpfully “normalizes” labels—for example, changing sec:conv to sec:convergence. You then cannot find the location by that label. To check, pick a few labels and search for their exact spelling in the source.
Step 2: Extract a paper-wide glossary of terms and symbols
Mathematical symbols are strong links between sections, but they are also easy for AI to interpret through familiar conventions. For example, \lrate might mean a fixed step size throughout your paper, even though it is commonly associated with a learning rate elsewhere.
Ask for a table with three columns: source notation, including the macro name | meaning in this paper | first occurrence, by section or equation number. Explicitly say: “Use this table as the reference for our subsequent conversation.”
You want a global dictionary that spans the paper and can be consulted for every later question, rather than a few temporary definitions provided while explaining one equation.
The following standalone teaching example contains symbol macros, equation labels, and cross-section references. Put it in the main file of an EasyLaTeX test project, compile it, and check whether AI can extract a complete glossary. Its prose only illustrates structure and references; it is not a real convergence proof or a result established by a research paper.
\documentclass[11pt]{article}
\usepackage{amsmath}
\usepackage{natbib}
\newcommand{\lrate}{\eta}
\newcommand{\loss}{J}
\begin{document}
\section{Introduction}
\label{sec:intro}
We study how a fixed step size $\lrate$ affects
convergence on $\loss$.
\section{Method}
\label{sec:method}
Let $\lrate$ be the step size and $\loss(\theta)$ the
objective. The update rule is Equation~\eqref{eq:update}.
\subsection{Update rule}
\label{sec:update}
\begin{equation}
\theta_{t+1} = \theta_t - \lrate \nabla \loss(\theta_t)
\label{eq:update}
\end{equation}
\section{Convergence}
\label{sec:conv}
Under Assumption A, we bound the number of steps
needed to reach a stationary point.
\section{Experiments}
\label{sec:exp}
Section~\ref{sec:conv} predicts a slower rate under a
decaying $\lrate$; we test this claim here.
\begin{thebibliography}{9}
\bibitem[Bottou(2012)]{bottou2012}
Bottou L. Stochastic Gradient Descent Tricks. 2012.
\end{thebibliography}
\end{document}
Important: Preserve macro names exactly in the table. \lrate may expand to the displayed symbol \eta, but they are different source expressions. Recording only the displayed symbol or a verbal definition loses the macro name and makes precise searches and consistent edits harder. Keep the source macro, displayed symbol, and actual meaning together.
Step 3: Ask for the argument, then verify its cited locations
Ask AI to summarize the paper as problem → method → experiments → conclusions, using one or two sentences for each part. Each sentence should identify the section, equation, or table supporting it.
Specific locations make misunderstandings visible. If it says, “The experiments in Section 4 validate the assumptions in Section 2,” checking the section numbers lets you see whether that relationship is correct.
Follow its references back to the source. If they do not match, first confirm that the relevant files were read and no sections were omitted. Then distinguish an AI misunderstanding from unclear wording in the paper itself. Record anything that remains unclear after verification for further investigation.
Step 4: Use the index to locate passages, and reread the source for detailed edits
After the first three steps, name index entries directly. For example: “Read the file containing sec:conv and the experimental section. Use the bound there to analyze why the third configuration in Section 4 converges slowly, and cite the evidence.” The index helps AI locate material; it does not replace that material. For detailed rewriting, ask the project AI to reread the relevant file and adjacent context before giving edit instructions. Supply the source yourself only when the material is not already available in the current project or conversation.
Keep a copy of the index you have checked. You can refer to it while continuing the conversation, but do not assume AI will remember every earlier detail indefinitely. Long conversations may undergo context compression. After switching projects or clearing history, provide the index again and confirm which materials are available. Regenerate it after substantial changes to the paper, explicitly retiring the previous version.
Multi-file projects: trace inclusion order from the compilation main file
If sections are split across files, make sure the relevant source is in the current EasyLaTeX project and check the project's compilation main file. Ask AI to start there, follow \input and \include in order, and read the included chapters before building the index. Summarizing only the inclusion commands in main.tex is not enough. Missing files should be reported explicitly, not filled in by guessing. You do not need to copy the entire project into the chat on every turn.
Summary
In the EasyLaTeX editor, try this workflow:
- Index first: Ask for sections and original
\labelvalues to establish shared reference points. - Put symbols in a table: Keep macro names and meanings together to prevent misinterpretation and inaccurate edits.
- Make the argument verifiable: Require citations. If they do not match, check reading coverage before judging the interpretation or source wording.
- Ask by index entry: Point to the relevant material in each follow-up instead of repeatedly pasting the full paper.