Academic Paper Analyzer & Figure Planner

Academic Paper Analyzer & Figure Planner

Produce an executable Figure Plan. No palette tables here — hand venue/domain/figure types to color-expert later.

Missing info: → references/missing-info-policy.md

Input Contract

• Prefer: paper PDF/LaTeX/Word, section drafts, abstract, method/experiments, repo quick-understanding doc, extracted architecture notes
• Minimum: title+abstract, or one method/experiment section, or a repo understanding doc
• Missing: partial plan with 推断 / 待确认

Output Contract — Figure Plan

• paper overview (topic, contributions)
• completeness block
• per-section figure recommendations
• priority ranking (must / strong / nice)
• palette: style family hint (classic vs pastel) + venue: + domain: — not hex tables; see references/palettes.md
• module_count_framework: (carried from repo Handoff when present)

Steps

Step 1: Parse structure

Map sections: Intro, Method (+ sub), Experiments, Analysis. Note missing sections.

Done when: section list exists and each is marked present / absent / partial.

Step 2: Mark figure-worthy content

content	figure type	priority
end-to-end pipeline	Overall Framework	must
network / layer structure	Network Architecture	must
novel module / mechanism	Module Detail	must
method variants / baselines	Comparison / Ablation	strong
representation / attention behavior	Data Behavior	strong / medium
dense math or loss	Module Detail	strong
curves / t-SNE / heatmaps	Data Behavior	medium

Done when: every must-level contribution has at least one figure entry or an explicit “insufficient evidence” note.

Step 3: Count and prioritize

paper class	typical count
top-conference long	6–8
short / workshop	4–5
journal	8–12
arXiv tech report	5–7 flexible

Done when: total count + must/strong/nice table is filled.

Step 4: Emit Figure Plan report

Include per-section: type × count, why, must-appear visual elements, aspect ratio hint.

type	aspect	core elements
Overall Framework	16:9	input → stages → output; innovation callouts
Network Architecture	16:9 / 3:2	layers, dims, residuals
Module Detail	4:3	central mechanism, ops (⊗ ⊕ σ), sparse formula
Comparison / Ablation	16:9	N×M grid, ours highlighted
Data Behavior	4:3 / 1:1	multi-panel heatmaps / curves / embeddings

Emit aspect_ratio per figure in the Figure Plan so prompt skills can copy it into the spec.

Done when: report matches Output Contract and completeness block is honest.

Domain packs (optional cues)

• CV: framework + arch + module + visual comparison + attention maps
• NLP: framework + transformer arch + attention module + metrics + embeddings
• RL/Robotics: state→policy→action loop + networks + trajectories
• Medical: imaging pipeline + U-Net/ViT + qualitative grid + ROC/features
• Protein / AI4Science: folding pipeline + Evoformer/module detail + MSA or structure behavior
• GNN / Scientific computing: message-passing module + encoder-processor-decoder framework + rollout trajectories

Sparse-input cases

materials	plan style
title + abstract only	high-level types only; no fake submodules
method without experiments	plan method figures; results as placeholders
partial sections	local plan; separate covered vs uncovered
only repo understanding doc	system-centric draft; flag narrative review needed

Stop

Stop when the Figure Plan for available materials is delivered. Do not generate prompts unless the user asks.