Di Tang

Di Tang

Associate Professor, SYSU

Paper News Generation Workflow

Paper News Generation Workflow

Use this workflow when generating _paper_news/YYYY-MM-DD.md from a daily AI paper report.

Goal

The page should read like a research brief, not a pipeline log. The first screen must answer:

  • What is the shortest useful headline for today’s research movement?
  • What are the 3 focus areas a researcher can understand at a glance?
  • Which paper should a researcher open first?
  • What is the biggest shift or warning from today’s papers?

The generated Markdown has two jobs:

  1. Preserve the full daily report body for deep reading.
  2. Add structured front matter that the Jekyll layouts use to create the visual top section on /paper-news/ and /paper-news/YYYY-MM-DD/.

Current Layout Contract

The current site uses these files:

  • _pages/paper-news.md: the index page at /paper-news/.
  • _layouts/paper-news-index.html: removes the default page title so the research brief is the first visual element.
  • _layouts/paper-news.html: renders an individual daily issue.
  • assets/css/main.scss: styles the researcher-focused paper-news components.

The index page uses the latest English issue’s front matter:

  • brief_title and brief_dek become the main headline.
  • themes[0..2] become the three focus cards under the headline.
  • lead_paper becomes the Best first paper item in Start here.
  • top_papers[1] becomes Also worth opening.
  • signals[0] becomes Biggest shift.
  • brief_dek is also used in the recent-issues list.

The daily issue page uses:

  • brief_title, brief_dek
  • takeaways
  • themes
  • lead_paper
  • signals
  • top_papers

If any field is missing, the page will still build, but the top section will feel empty. Do not omit required fields.

Extraction Recipe From The Raw Daily Report

Start from the raw daily report body, usually containing:

  • Executive takeaways
  • Key themes (clusters)
  • Technical synthesis
  • Top 5 papers
  • selected-paper table with arXiv IDs, scores, reasons, and tags

Generate the structured front matter as follows:

brief_title

Extract the dominant movement across the day. Use 3-6 words. It should be a claim, not a label.

Good:

  • AI reliability gets real.
  • Agent evaluation gets harder.
  • Robustness moves downstream.

Bad:

  • Daily AI Paper Report
  • Today’s AI Papers
  • What should I read?

brief_dek

Condense the top 2-3 executive takeaways into one sentence. It should explain why the title is true.

themes

Create exactly 3 themes from Key themes (clusters) or the strongest recurring ideas in Executive takeaways.

Each theme must be short enough for a card:

  • title: 2-4 words.
  • text: one concrete sentence, preferably under 22 words.

The index page labels these cards as:

  • Why it matters
  • What changed
  • What to watch

So order the themes accordingly:

  1. The most important implication.
  2. The clearest shift from prior expectations.
  3. The thing to monitor next.

takeaways

Create exactly 3 high-level conclusions from Executive takeaways. These can be longer than theme card text because they are shown on the daily issue page, not the compact index hero.

lead_paper

Choose the paper a researcher should open first. Do not pick purely by score. Prefer a paper with:

  • a reusable method,
  • a surprising result,
  • strong deployment relevance,
  • a claim likely to be cited,
  • or a useful limitation.

Fill:

  • title
  • url
  • pdf
  • why: one sentence explaining why this paper should be opened first.
  • skepticism: one sentence with the main caveat.
  • tags: 2-4 short tags.

top_papers

Create 3-5 entries from the Top 5 papers section. The first item should normally match lead_paper.

For each:

  • why: why a researcher should spend time on it.
  • why_now: why it is timely now.
  • skepticism: what could limit the claim.

The index page uses top_papers[1] as Also worth opening, so make the second paper genuinely complementary to the first.

signals

Create exactly 3 signals. These are not topics; they are research movements, warnings, or method patterns.

The index page only shows signals[0] as Biggest shift, so put the strongest signal first.

Each signal:

  • label: concrete category, not generic wording.
  • title: specific claim a researcher can agree or disagree with.
  • text: concrete evidence from the day’s papers.

Good labels:

  • Evaluation shift
  • Agent pattern
  • Robustness bet
  • Safety warning
  • Deployment gap
  • Method pattern

Avoid labels like:

  • Signal
  • Tension
  • Bet
  • Theme
  • Insight

Length Budget For The Visual Top Section

The front matter controls a compact visual layout. Keep it tight:

  • brief_title: 3-6 words.
  • brief_dek: 18-35 words.
  • themes.*.title: 2-4 words.
  • themes.*.text: under 22 words.
  • signals.*.title: under 9 words.
  • signals.*.text: under 28 words.
  • lead_paper.why: under 30 words.
  • lead_paper.skepticism: under 26 words.

Required File Path

Create the English issue at:

_paper_news/YYYY-MM-DD.md

Create the Chinese issue at:

_paper_news/YYYY-MM-DD-zh.md

Required Front Matter

Every new English issue should use this front matter schema:

---
layout: paper-news
title: "Daily AI Paper Report (YYYY-MM-DD)"
date: YYYY-MM-DD
permalink: /paper-news/YYYY-MM-DD/
author_profile: false
brief_title: "Concise information-bearing headline."
brief_dek: "One sentence explaining the main research pattern and why it matters."
lead_paper:
  title: "Paper title"
  url: "https://arxiv.org/abs/..."
  pdf: "https://arxiv.org/pdf/....pdf"
  why: "Why this paper should be opened first."
  skepticism: "The most important limitation or caveat."
  tags:
    - short tag
    - short tag
takeaways:
  - "Most important research-level conclusion."
  - "Second important conclusion."
  - "Third important conclusion."
themes:
  - title: "Short theme name"
    text: "One sentence explaining the theme."
  - title: "Short theme name"
    text: "One sentence explaining the theme."
  - title: "Short theme name"
    text: "One sentence explaining the theme."
signals:
  - label: "Concrete category"
    title: "Specific research claim."
    text: "Concrete evidence or implication in one sentence."
  - label: "Concrete category"
    title: "Specific research claim."
    text: "Concrete evidence or implication in one sentence."
  - label: "Concrete category"
    title: "Specific research claim."
    text: "Concrete evidence or implication in one sentence."
top_papers:
  - title: "Paper title"
    url: "https://arxiv.org/abs/..."
    pdf: "https://arxiv.org/pdf/....pdf"
    why: "Why a researcher should spend time on it."
    why_now: "Why this is timely."
    skepticism: "Main caveat."
---

Writing Rules

  • brief_title should be short and information-bearing. Aim for 3-6 words.
  • Avoid vague titles like “What should I read today?” or “Today’s AI paper digest.”
  • Good title pattern: “AI reliability gets real.”
  • brief_dek carries the nuance. Keep it to one sentence.
  • lead_paper.why should describe the paper’s research value, not its score.
  • lead_paper.skepticism is mandatory; researchers trust pages that show judgment.
  • takeaways is mandatory. Use 3 high-level conclusions from the executive takeaways.
  • themes is mandatory. Use 3 theme cards that summarize the day’s intellectual clusters.
  • signals should summarize the day’s intellectual structure, not paper categories.
  • Avoid generic signal labels such as Signal, Tension, Bet, Theme, or Insight.
  • Use concrete labels that name the actual research movement, such as Evaluation shift, Agent pattern, Robustness bet, Safety warning, Deployment gap, or Method pattern.
  • Each signal title should make a specific claim a researcher can agree or disagree with.
  • Each signal text should mention concrete evidence from the day’s papers, not abstract phrasing.
  • top_papers should contain 3-5 papers. The template currently displays the first 3.
  • Keep telemetry such as candidates, selected count, and deepread completion inside the report body, not the front-page headline.

Body Order

After the front matter, keep the current long report body format:

  1. Language link.
  2. Optional run stats/details block.
  3. # AI Paper Insight Brief.
  4. Executive takeaways.
  5. Key themes.
  6. Technical synthesis.
  7. Top papers.

The paper-news layout renders the structured front matter first, then renders the full Markdown report below it.

Generation Checklist

Before saving the file, verify:

  • The front matter is valid YAML.
  • layout is paper-news.
  • brief_title is not the same as title.
  • themes has exactly 3 entries and each entry is visually short.
  • takeaways has exactly 3 entries.
  • signals has exactly 3 entries and signals[0] is the strongest because it appears on the index page.
  • lead_paper is not selected only by score; it is selected for research value.
  • top_papers has at least 3 entries and top_papers[1] is a good companion paper.
  • The report body still includes the full detailed analysis after the front matter.
  • Pipeline telemetry stays in the body, not in brief_title, brief_dek, themes, or signals.

Chinese Pages

Chinese pages can keep the existing layout: single until the Chinese brief schema is added. If adding a Chinese structured version, use translated field values and layout: paper-news.