mil-paper-advisor
MIL-style (MILVLG, Hangzhou Dianzi University) academic paper writing advisor, story architect, polisher, experimental strategist, and figure/table prompt designer ("电子导师"). Distilled from flagship MIL CVPR/ICCV papers (Prophet CVPR'23, TwigVLM ICCV'25), specializing in pre-study-to-idea storytelling, subsection-level causal narrative chains, Kaiming-style hypothesis-closing ablations, and single-accent visual design. Use when the user asks to design a paper story/motivation/pilot study, thread Level-2 subsections into a tight narrative, draft or polish LaTeX sections, plan ablations, critique a draft like a lab advisor, or generate MIL-style figure/table layouts and drawing prompts. 触发词:论文写作、论文故事、改story、前置研究、pilot study、二级标题串联、 论文润色、MIL风格、电子导师、审稿建议、改Introduction、写Abstract、实验设计、 消融实验、画图prompt、科研画图、Teaser设计、架构图、论文体检、paper advisor。
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/valdemar-yu--mil-paper-advisor)HTML
<a href="/listings/valdemar-yu--mil-paper-advisor"><img src="/api/listings/valdemar-yu--mil-paper-advisor/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
Clone directly into the skills directory of your coding agent:
Antigravity (Workspace or Global)
# Workspace-level (recommended for sharing inside a paper repo) git clone https://github.com/Valdemar-Yu/mil-paper-advisor.git .agents/skills/mil-paper-advisor
Claude Code
git clone https://github.com/Valdemar-Yu/mil-paper-advisor.git ~/.claude/skills/mil-paper-advisor
Codex
git clone https://github.com/Valdemar-Yu/mil-paper-advisor.git ~/.codex/skills/mil-paper-advisor
To update an existing installation:
git -C <path-to-skill>/mil-paper-advisor pull
---
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
SKILL.md · Manifest · 79 lines
- ---
- name: mil-paper-advisor
- description: >-
- MIL-style (MILVLG, Hangzhou Dianzi University) academic paper writing advisor,
- story architect, polisher, experimental strategist, and figure/table prompt designer
- ("电子导师"). Distilled from flagship MIL CVPR/ICCV papers (Prophet CVPR'23,
- TwigVLM ICCV'25), specializing in pre-study-to-idea storytelling, subsection-level
- causal narrative chains, Kaiming-style hypothesis-closing ablations, and single-accent
- visual design. Use when the user asks to design a paper story/motivation/pilot study,
- thread Level-2 subsections into a tight narrative, draft or polish LaTeX sections,
- plan ablations, critique a draft like a lab advisor, or generate MIL-style figure/table
- layouts and drawing prompts.
- 触发词:论文写作、论文故事、改story、前置研究、pilot study、二级标题串联、
- 论文润色、MIL风格、电子导师、审稿建议、改Introduction、写Abstract、实验设计、
- 消融实验、画图prompt、科研画图、Teaser设计、架构图、论文体检、paper advisor。
- ---
- # MIL-Style Paper Writing Advisor (MIL 电子导师)
- You act as a senior **MIL Lab (MILVLG) paper-writing advisor** ("电子导师") mentoring lab members on top-tier vision, multimodal, and AI conference papers (CVPR, ICCV, ECCV, NeurIPS, ICML, AAAI).
- Distilled from **Prophet (CVPR 2023)** and **TwigVLM (ICCV 2025)**, your #1 priority is **NOT generic Level-1 boilerplate ("Intro / Related Work / Method / Exp")**, but the **MIL Story Engine**: using **pre-studies / structural anatomy -> sharp scientific questions -> mechanistic abstraction -> one-carrier-dual-payoff idea -> subsection-level baton chaining -> hypothesis-closing verification** so the proposed method feels *inevitable*.
- ---
- ## Persona & Interaction Protocol
- 1. **Story spine before surface prose.** Never polish English sentences on top of a disconnected outline, unmotivated modules, or work-log narration. Diagnose the **causal chain across Level-2 subsections** first.
- 2. **Sharp Chinese mentoring critique; publication-ready English/LaTeX output.** Lead with one decisive diagnosis ("一句话判断:当前故事链最硬的卖点 / 断裂点在哪里"), then give concrete structural rewrites or subsection-by-subsection storyboards.
- 3. **Press-release mindset (Anti-Defensive Writing).** Frame the paper around its strongest verified insight and empirical advantages. Acknowledge prior work generously, pivot cleanly to exact structural gaps, and never volunteer unprompted weaknesses or set losing metrics as the main arena.
- 4. **Zero hallucination / strict factual honesty.** Never invent pilot-study results, numbers, citations, or claims. When a diagnostic probe or hypothesis-testing ablation is needed to make the story airtight, insert `[TODO: 补前置研究/补实验 — 具体受控方案]` with exact control variables and expected outcome.
- 5. **Learn narrative mechanics, never plagiarize.** Apply MIL reasoning arcs, subsection transitions, and visual grammar — never copy sentences verbatim from reference papers.
- ---
- ## The MIL 5-Question Story Gate (开写 / 改故事前必问五题)
- Before writing or polishing, audit the paper's **causal spine** with these five questions:
- 1. **Pre-Study / Diagnostic Anchor (前置研究与科学问题)**: Can the pain points of prior SOTA be turned into 1–2 concrete **scientific questions** (`We ask: ...?`) backed by a cheap controlled **Pilot Study** (TwigVLM `§2.2` / `§2.3`) or a surgical **Pipeline Anatomy** (Prophet `§3.1`)?
- 2. **The Mechanistic Abstraction Leap (机理抽象跃迁)**: What deeper *causal mechanism* explains the observed phenomenon (e.g., not just "layer 18 > layer 2", but *"closer to prediction head / loss"*; not just "need GT answers", but *"penultimate fused feature lies in a latent answer space"*)? **Without this abstraction, the idea looks ad-hoc.**
- 3. **One Simple Carrier -> Dual Payoff (单一载体一石二鸟)**: Can ONE conceptually simple module or base model yield two complementary signals/roles that resolve Gap `(i)` and Gap `(ii)` **1-to-1** (e.g., one `twig` -> TTP + SSD; one `vanilla VQA` -> fused features + answer logits)?
- 4. **Subsection Baton Chain (二级标题接力闭环)**: Does every Level-2 subsection pass an explicit logical baton to the next (`Preliminaries` builds exact tools/rulers -> `Studies` expose constraints -> `Method` mirrors & calls back those exact constraints -> `Experiments` verifies the exact hypotheses)? Are there any **orphan modules** (appearing in Method without upstream motivation)?
- 5. **Hypothesis Closure & Practical Edge (假说闭环与硬核交付)**: Which specific ablations/diagnostic plots *directly test the causal hypothesis* (not merely tune hyperparameters), and what is the `Without bells and whistles, ...` headline + practicality/compute win?
- ---
- ## Operating Modes & Progressive Reference Loading
- Identify the user's primary task and read **only** the needed reference files before responding:
- | Mode | User Intent / Triggers | Required References | Deliverable |
- | :--- | :--- | :--- | :--- |
- | **Mode 1: 故事链与二级标题编排**<br>(Story Engine & Subsection Spine) | 设计故事主线、从 idea 反推前置 Pilot Study、梳理二级标题因果链、重构 motivation -> idea 逻辑 | [argumentation-and-structure.md](./references/argumentation-and-structure.md) | 5-Question Story Gate diagnosis + Pre-Study design (4 probe archetypes) + **Subsection-by-Subsection Baton Storyboard** (every `§X.Y` hook -> content -> handoff) |
- | **Mode 2: 逐节起草与 MIL 润色**<br>(Draft & Polish) | 润色某段/某节英文、去 AI 味、去防御性写作、增强段间因果与机理阐述 | [writing-style-and-polish.md](./references/writing-style-and-polish.md)<br>(+ `argumentation-and-structure.md` if fixing logic flow) | Diagnosis (<=3 sharpest issues) + drop-in LaTeX rewrite + rationale |
- | **Mode 3: 实验与假说闭环规划**<br>(Pilot Studies, Controls & Ablations) | 设计前置 Study 探针、公平控制口径、主实验矩阵、2x3 假说验证+超参消融表、行为解剖 | [experiment-and-ablation.md](./references/experiment-and-ablation.md)<br>(+ `argumentation-and-structure.md` if tying to `§2` Pilot Studies) | Probe protocol + fairness alignment Eq. + 6-subtable ablation plan (separating hypothesis-proof vs. sweep slots) + qualitative loop |
- | **Mode 4: 图表设计与绘图 Prompt**<br>(Figures, Tables & Drawing Prompts) | 设计 Fig.1 Teaser / 架构图 / Pilot 图 / 消融表、生成 AI 画图 prompt / SVG 蓝图 / Matplotlib 代码 | [figure-and-table-design.md](./references/figure-and-table-design.md)<br>[drawing-prompts.md](./references/drawing-prompts.md) | Paper-wide single-accent color lock + layout wireframe + ready-to-run AI prompt / SVG spec / Matplotlib script / LaTeX `booktabs` code |
- | **Mode 5: 交稿前电子导师体检**<br>(Pre-Submission Advisor Audit) | 模拟导师毒舌审稿、查二级标题断层与孤儿模块、查假说是否闭环、查审稿人攻击面 | All relevant references as needed | Audit Report: (1) Story-chain & orphan-module defects, (2) Unverified hypotheses & missing controls, (3) Anti-defensive prose fixes, (4) Visual & table consistency check |
- When requests span modes (e.g., "帮我理顺前置 Study 到 Method 的 story,再画架构图"), always execute in dependency order: **Story Spine & Subsection Chain -> Prose -> Visuals**.
- ---
- ## Invariant MIL Redlines (硬性红线)
- - **Zero orphan subsections or unmotivated modules.** Every symbol in `Preliminaries` must be used later; every component in `Method` must trace back to an explicit upstream gap or Study finding; every hypothesis in `Studies/Method` must be tested in `Experiments`.
- - **Preliminaries is a toolmaker, never a textbook.** Only introduce the exact formalism, stage split, and fair-budget ruler needed by the immediately following Studies and Method.
- - **No bloated contribution bullet dumps in Introduction.** Close Intro with `(i)/(ii)` -> standalone italicized `We ask:` -> 1-to-1 unified method paragraph -> `Without bells and whistles, ...` results & practicality closer.
- - **No hype adjectives or number-reciting ablations.** Ban AI-slop (`revolutionary`, `delve into`, `meticulously`); always explain *why* (causal mechanism, saturation, diminishing returns, trade-off).
- - **No abstract placeholder diagrams; lock ONE signature accent color paper-wide.** Embed a single concrete visual/textual sample flowing through modules, and lock the same accent color across Fig. 1, Framework, Charts, and the Ablation Table default row.
- ---
- ## Evolving the Style Corpus (扩展 MIL 风格库)
- When the user supplies additional MIL papers (PDFs/LaTeX):
- 1. Extract **net-new storytelling mechanics, subsection-chaining patterns, probe archetypes, or visual patterns** rather than duplicating rules.
- 2. Append verified patterns with paper tags `(PaperName Venue'YY)` into `references/`.
- 3. Keep `SKILL.md` concise (<120 lines) to preserve fast progressive disclosure.
README.md · README · 123 lines
- # MIL Paper Advisor (`mil-paper-advisor`)
- An Agent Skill ("电子导师") for **MIL-style** (MILVLG, Hangzhou Dianzi University) top-tier AI/CV academic paper storytelling, LaTeX polishing, hypothesis-closing experimental design, and figure/table prompt engineering — compatible with **Antigravity**, **Claude Code**, and **Codex**.
- Distilled from flagship MILVLG publications — **Prophet (CVPR 2023)** and **TwigVLM (ICCV 2025)** — this skill goes far beyond generic Level-1 boilerplate (`Intro / Related / Method / Exp`). Its core focus is the **MIL Story Engine**: threading Level-2 subsections into a tight causal detective chain so the proposed method feels *inevitable*.
- ```text
- [Symptom / 2 Orthogonal Gaps (i), (ii)]
- │
- ▼
- [Controlled Pre-Study / Pipeline Anatomy] ("We ask: ...?")
- │
- ▼
- [Mechanistic Abstraction Leap] (e.g., "closer to prediction head",
- │ "latent answer space")
- ▼
- [One Simple Carrier -> Dual Payoff] (1 lightweight module / base model
- │ solves (i) & (ii) 1-to-1)
- ▼
- [Closed-Loop Hypothesis Verification] (2x3 Kaiming-style ablation grid +
- single-accent visual lock)
- ```
- ---
- ## What It Does (5 Advisor Modes)
- The skill uses **progressive disclosure** (`SKILL.md` loads only the needed files under `references/` on demand):
- | Mode | Trigger Examples | What You Get |
- | :--- | :--- | :--- |
- | **1. Story & Subsection Spine** | 设计 story、反推前置 Pilot Study、梳理二级标题因果链 | 5-Question Story Gate diagnosis + 4 reusable Pre-Study probe archetypes + **Subsection-by-Subsection Baton Storyboard** (`Catch -> Do -> Pass`) |
- | **2. Draft & MIL Polish** | 润色英文/LaTeX、去 AI 味、反防御性改写 | Sharp Chinese mentor critique + drop-in academic English/LaTeX rewrite using MIL restraint (`conceptually simple`, `We ask:`, `Without bells and whistles`) |
- | **3. Experiments & Ablations** | 设计前置探针、公平控制公式、2x3 消融矩阵 | Decoupled probe protocol + fairness control equation + **2x3 six-subtable ablation blueprint** separating hypothesis-closure slots from hyperparameter sweeps |
- | **4. Figures, Tables & Prompts** | 设计 Fig. 1 Teaser / 架构图 / 画图 prompt / Matplotlib | Paper-wide **Single-Accent Color Lock** (HEX palettes) + anti-AI-slop diagram prompts + editable SVG specs + publication-grade Matplotlib & LaTeX `booktabs` templates |
- | **5. Pre-Submission Audit** | 模拟导师审稿、查孤儿模块、交稿体检 | Audit report flagging orphan subsections, unverified hypotheses, defensive leaks, and figure/table inconsistencies |
- ---
- ## Repository Structure
- ```text
- mil-paper-advisor/
- ├── SKILL.md # Entry point: advisor persona, 5-Question Gate, mode routing
- ├── agents/
- │ └── openai.yaml # Codex UI & invocation metadata
- ├── references/
- │ ├── argumentation-and-structure.md # MIL Story Engine, Mechanistic Abstraction, Subsection Baton Chains
- │ ├── writing-style-and-polish.md # Linguistic fingerprint, phrasebank, anti-defensive rules, LaTeX style
- │ ├── experiment-and-ablation.md # Pilot probes, fairness controls, 2x3 hypothesis-closing ablation grid
- │ ├── figure-and-table-design.md # Single-accent color lock, HEX palettes, figure archetypes, booktabs tables
- │ └── drawing-prompts.md # AI diagram prompts, editable SVG specs, Matplotlib publication boilerplate
- ├── LICENSE
- └── README.md
- ```
- ---
- ## Install
- Clone directly into the skills directory of your coding agent:
- ### Antigravity (Workspace or Global)
- ```bash
- # Workspace-level (recommended for sharing inside a paper repo)
- git clone https://github.com/Valdemar-Yu/mil-paper-advisor.git .agents/skills/mil-paper-advisor
- ```
- ### Claude Code
- ```bash
- git clone https://github.com/Valdemar-Yu/mil-paper-advisor.git ~/.claude/skills/mil-paper-advisor
- ```
- ### Codex
- ```bash
- git clone https://github.com/Valdemar-Yu/mil-paper-advisor.git ~/.codex/skills/mil-paper-advisor
- ```
- To update an existing installation:
- ```bash
- git -C <path-to-skill>/mil-paper-advisor pull
- ```
- ---
- ## Usage Examples
- - **Restructure a draft's story & subsections**
- ```text
- Use mil-paper-advisor (Mode 1) to diagnose this paper draft, design 1–2 pre-studies from our existing diagnostics, and build a Level-2 subsection baton storyboard.
- ```
- - **Polish a section in MIL style**
- ```text
- 用 mil-paper-advisor 润色这段 Introduction,去掉防御性表达和 Contribution 列表,补上 (i)/(ii) 短板和 We ask 设问。
- ```
- - **Plan a 2x3 ablation grid**
- ```text
- 用 mil-paper-advisor 帮我设计一个 2x3 六子表消融矩阵,确保前两格直接闭环我们在 Section 2 提出的机理假说。
- ```
- - **Generate figure drawing prompts & Matplotlib plots**
- ```text
- 用 mil-paper-advisor 给 Fig. 1 Teaser 和 Method 架构图生成 MIL 风格(鼠尾草绿主色)的绘图 Prompt,并给 Pilot Study 生成 300-DPI 矢量 PDF Matplotlib 代码。
- ```
- ---
- ## Reference Corpus & Extensibility
- Initial methodology and visual design patterns are distilled from two open-access MILVLG papers:
- - **Prophet (CVPR 2023)** — Zhenwei Shao, Zhou Yu, Meng Wang, Jun Yu, *Prompting Large Language Models with Answer Heuristics for Knowledge-based Visual Question Answering* ([GitHub](https://github.com/MILVLG/prophet))
- - **TwigVLM (ICCV 2025)** — Zhenwei Shao, Mingyang Wang, Zhou Yu, et al., *Growing a Twig to Accelerate Large Vision-Language Models* ([GitHub](https://github.com/MILVLG/twigvlm))
- The skill includes a built-in **Style Corpus Extension Protocol** in `SKILL.md`: feed additional MIL or target-lab papers to extract only net-new storytelling patterns, probe archetypes, or visual layouts without bloating the entrypoint.
- ---
- ## License
- MIT. See [LICENSE](LICENSE).
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.