Longform Paper Review Guide

Use this guide before writing a public paper review that should read like a blog post rather than a compact wiki note. The target is a beginner-friendly review that a reader can finish in about 20-30 minutes while still leaving enough structure for later wiki reuse.

This guide is especially appropriate for foundational papers such as architecture, learning-method, generative-model, evaluation, or computational biology papers.

Review Goal

A longform paper review should answer one reader question:

$$\text{paper}\to \text{problem}\to \text{idea}\to \text{method}\to \text{evidence}\to \text{limits}\to \text{why it matters}$$

It should not be a section-by-section translation of the paper. It should explain why the paper was needed, what changed, how the method works, what evidence supports it, and which concepts the reader should learn next.

Length and Reading Budget

Target	Guideline
Reading time	20-30 minutes
Main sections	7-10 sections
Math depth	enough to understand the central claim
Tables	use for comparison, routing, evidence, limitations
Figures	use self-made diagrams, Mermaid diagrams, or concise schematic descriptions
External claims	cite or mark to verify; do not invent metadata, metrics, or results

If the review becomes longer than this, split reusable definitions into Concepts, formulas into Math, and evidence details into Paper analysis.

Required Structure

Use this shape unless the paper strongly suggests a better one.

Section	Purpose	Typical Links
One-line summary	state the paper’s core contribution in plain language	paper bucket, main concept
Why this paper mattered	explain the problem before the method	AI, Math, Computational Biology when relevant
Background before the paper	define the baseline or older paradigm	architecture, learning, evaluation, modality notes
Paper’s main idea	name the key move and what it replaces	concept notes and formulas
Method walkthrough	explain components in dependency order	equations, diagrams, route tables
Experiments and evidence	connect claims to tables, benchmarks, baselines, and ablations	Evidence table, Benchmark card
What the paper does not prove	limits, assumptions, missing comparisons, leakage or scaling caveats	Limitation taxonomy
Concept map	list the wiki pages the reader should open next	AI, Math, Computational Biology, Infra, Agents
Takeaways	explain what should remain in memory after reading	related papers or project ideas
Open questions	mark uncertainty, reproduction targets, or follow-up reading	reproducibility and research-method notes

Beginner-Friendly Writing Rules

• Start with the problem, not the model name.
• Define every important object before using equations.
• Explain what each equation computes and why it appears.
• Prefer the paper’s causal story over a chronological summary.
• Use short paragraphs and descriptive section titles.
• Put comparisons into tables rather than long prose.
• Use diagrams when several tensors, modules, or evidence paths interact.
• Do not paste copyrighted figures from the paper. Recreate minimal schematic diagrams or describe the figure’s role.
• Do not copy long passages from the paper. Paraphrase and link the public source.
• Mark unknown metadata, artifact status, and uncertain claims as to verify.

Formula Depth

Use Formula explanation ladder before writing formulas.

Case	Depth
Formula is background only	define the intuition and link to Math
Formula is the paper’s main contribution	include the full canonical equation and symbol definitions
Formula explains a model block	include input, intermediate variables, output, and shape or dimension when useful
Formula explains evaluation	include metric definition, denominator, split, and what higher/lower means

For architecture papers, avoid shorthand. For attention, define:

$$Q=X{W}_Q,K=X{W}_K,V=X{W}_V$$ $$\mathrm{Attention}(Q,K,V)=\mathrm{softmax}\left(\frac{Q{K}^{\top }}{\sqrt{d_k}}\right)V$$

Then explain what $Q$, $K$, $V$, $d_k$, softmax, masking, heads, and output projection do. If a causal or padding mask is used, include the mask term:

$$\mathrm{softmax}\left(\frac{Q{K}^{\top }}{\sqrt{d_k}}+M\right)V$$

Visual and Table Policy

Use visuals to reduce cognitive load, not to decorate.

Need	Preferred Form
Compare old vs new method	two-column table
Explain a pipeline	Mermaid flowchart
Explain architecture blocks	compact schematic or ordered block table
Explain experiment claims	claim-evidence table
Explain benchmark limits	benchmark card
Explain formulas	equation plus symbol table

For longform reviews, at least one of these should usually appear:

• problem-to-method table;
• method component table;
• formula symbol table;
• evidence table;
• limitation table;
• concept map.

Wiki Link Policy

Every longform paper review should include links that help the reader leave the post when they need prerequisites.

Link Type	Use
AI	architecture, learning method, generative model, evaluation, systems
Math	probability, linear algebra, optimization, information theory, geometry, metrics
Computational Biology	protein, molecule, ligand, structure, interaction, docking, data/evaluation
Concepts	reusable definitions and formulas
Papers	related papers and comparisons
Infra	compute, training, serving, reproducibility, systems claims
Agents	agent workflow or tool-use papers

Do not repeat full prerequisite tutorials inside the review. Link the canonical note, then explain only the minimum needed for this paper.

Evidence Discipline

Use the paper’s evidence only within the scope it supports.

Evidence Item	Must State
Dataset or benchmark	task, split, metric, allowed information
Result table	metric direction, baseline, selection rule, uncertainty when available
Ablation	what component was isolated and what remained controlled
Scaling result	parameter count, data, compute, sequence length, hardware or implementation boundary
Artifact	public code, data, configs, weights, logs, predictions, environment

When the paper does not provide enough information, write to verify rather than guessing.

Longform Review Skeleton

---
title: Paper Title
aliases:
  - papers/short-slug
tags:
  - papers
  - ai
---
 
# Paper Title
 
> One-line summary.
 
## Metadata
 
| Field | Value |
| --- | --- |
| Paper | to verify |
| Authors | to verify |
| Venue / year | to verify |
| Link | to verify |
| Status | reading |
 
## Why This Paper Mattered
 
Problem before the paper.
 
## Background
 
What a beginner must know first.
 
## Main Idea
 
The core move.
 
## Method Walkthrough
 
Component-by-component explanation.
 
## Core Formula
 
Equation, symbol table, and intuition.
 
## Experiments and Evidence
 
| Claim | Evidence | Caveat |
| --- | --- | --- |
| to verify | to verify | to verify |
 
## Limitations
 
What the paper does not prove.
 
## Concept Map
 
- [[ai/index|AI]]
- [[math/index|Math]]
- [[papers/index|Papers]]
 
## Takeaways
 
What to remember.
 
## Open Questions
 
What remains unresolved.

Pre-Publish Checklist

• Metadata is verified or marked to verify.
• The paper’s problem is clear before method details.
• The main contribution is stated in one sentence.
• The method section is ordered by dependency, not by paper section order.
• Central formulas include symbol definitions.
• At least one table or diagram improves readability.
• Claims are tied to evidence, baseline, metric, split, or ablation.
• Limitations and unsupported claims are explicit.
• AI, Math, Computational Biology, Infra, or Agent links are added where useful.
• No private project details, unpublished results, private paths, server details, or credentials appear.
• The note can be read as a public blog-style explanation in 20-30 minutes.

Related

• Papers
• Paper review workflow
• Paper note format
• AI Computational Biology Math paper template
• Claim routing
• Paper to wiki extraction
• Evidence table
• Benchmark card
• Ablation map
• Formula explanation ladder
• Coverage matrix