ICLR Blog Post Checklist

ICLR Blog Post Checklist

Write Research-Grade Posts That Engage Both Practitioners & Researchers

PRE-WRITING: The Foundation

Define Your Core Contribution

  • What is the novelty? Identify one non-obvious insight (algorithm insight, empirical finding, theoretical connection, or reframing)
  • Who cares and why? Map stakeholders: ML practitioners, a specific subfield, theoreticians, practitioners, or cross-disciplinary audiences?
  • What problem does this solve? Frame around pain points, not features
  • Why now? What recent work, trend, or limitation makes this relevant in 2026?
  • Is this ICLR-track worthy? Could this spark discussion at a venue? Does it advance the field’s thinking?

Research & Context

  • Identify 3–5 foundational papers (go to originals, not just surveys)
  • Map the lineage: How did we get here? What prior work built the path?
  • Find the gap: Where is the community’s blind spot, assumption, or limitation?
  • Gather empirical anchors: Specific numbers, benchmarks, results to ground claims
  • Review related blog posts (Distill, Anthropic, OpenAI, DeepMind) for tone and structure

STRUCTURE: Architect for Navigation & Depth

Opening Strategy: Start with a Concrete Hook

  • Lead with a scene, question, or surprising finding (not a definition or background dump)
    • ❌ “Attention mechanisms are a key component of transformers…”
    • ✅ “A 7B-parameter model trained on diverse data sometimes outperforms a 100B model on specialized tasks. Why?”
  • State the motivation in one sentence: Why should a busy researcher care?
  • Promise the payoff: What will they understand by the end?
  • Include reading time estimate at top (15–45 min for ICLR-level posts)

Build a Map Early

  • Create a Table of Contents with clear section hierarchy (H1 → H2 → H3)
    • Each section should have one purpose, defensible in a sentence
    • Section headings should be descriptive and searchable, not clever
    • Use numbered sections (1., 2., etc.) for navigability
  • Add a “Roadmap” paragraph after intro explaining the narrative flow
    • “We’ll first review X, then empirically test Y, then explore the implications for Z”
  • Use anchor links throughout so readers can jump to relevant sections

Middle Structure: Evidence → Explanation → Implication

  • Each major section opens with a concrete thing (result, diagram, theorem statement, code snippet)
  • Then explain: What does it mean? Why should we believe it? Where does it fit in the narrative?
  • Close with implication: How does this change our thinking? What’s the next question?

Subsection Discipline

  • Break long sections (>1000 words) into clearly labeled subsections (H3/H4)
  • Each subsection should be self-contained yet cohesive—readers can skip or revisit
  • Use bold topic sentences at paragraph starts to enable skim reading
  • Limit paragraphs to 4–5 sentences; one idea per paragraph

Closing: Relentless Summarization

  • Include a “Closing Thoughts” section that distills the entire post into 5–7 bullet points
  • Recap what changed in the reader’s mental model
  • Explicitly state limitations and open questions
  • Point to concrete next steps or follow-up work
  • Avoid soft conclusions; be direct about implications

WRITING: Rigor Meets Clarity

Lead with Intuition, Ground in Math

  • Explain the intuition first, then the formalism
    • Before: $$\mathcal{L} = \mathbb{E}[D_{KL}(q_\phi(z x) | p(z))] + \mathbb{E}[D_{KL}(p(x z) | q_\phi(x))]$$
    • After: “Variational autoencoders balance two competing goals: compress the input into a bottleneck (the KL term) while still reconstructing it accurately (the reconstruction term). Here’s why:” [then show formula]
  • Use thought experiments to build intuition
    • “Imagine training a 1B model on 1T tokens, then a 10B model on the same tokens. Which learns faster?”
  • Provide analogies grounded in familiar domains
    • Scaling laws ↔ returns on investment in engineering
    • Attention mechanisms ↔ selective focus in human vision
    • Latent representations ↔ abstract concepts in psychology

Explain the “Why” Chain

  • For each technical claim, answer: “Why is this true?” or “Why do we care?”
  • For algorithmic choices: “Why this design and not another?”
  • For empirical results: “What’s the underlying mechanism?”
  • For limitations: “What would it take to overcome this?”

Terminology & Definitions

  • Define technical terms at first use, not in a separate glossary
  • Use consistent terminology throughout (pick one term per concept; note synonyms once)
  • For domain-specific jargon, provide a brief parenthetical or footnote
  • If the field uses a term loosely, flag it: “What practitioners call ‘fine-tuning’ actually encompasses…”

Tone & Voice

  • Write conversationally but rigorously—no marketing hype, no filler
  • Use active voice where possible; passive only when emphasizing the action, not the actor
  • Be opinionated (flag opinions explicitly: “We believe…” or “Importantly…”)
  • Be honest about uncertainty: “This is speculative, but…” or “The evidence is mixed…”
  • Avoid generic transitions (“Furthermore,” “Moreover”); use specific connectors (“As we discussed earlier…” or “This contrasts with…”)

Avoid These Anti-Patterns

  • ❌ Generic openings: “In today’s rapidly evolving landscape of AI…”
  • ❌ Vague claims without evidence: “Attention is revolutionary”
  • ❌ Invented metrics or unsourced quotes
  • ❌ Passive constructions masking attribution: Use active voice and cite sources
  • ❌ Tangential depth: Keep main text focused; use footnotes or sidebars for asides
  • ❌ Overstated claims beyond what evidence supports

VISUALS: Strategic Diagrams & Figures

Figures Must Explain, Not Decorate

  • Every figure should answer a specific question
    • ❌ “Here’s what an attention head looks like”
    • ✅ “Attention heads specialize: head 1 attends locally, head 2 attends to specific tokens”
  • Include a descriptive caption for each figure: “Figure N: What this shows and why it matters”
  • Use consistent visual style across all figures (colors, fonts, line weights)
  • Ensure figures are readable at 80% zoom (critical for small screens)

Diagram Types for ICLR Posts

  • Architecture diagrams: Clearly show data flow and skip connections
  • Learning curves: Include error bars or confidence intervals; label convergence behavior
  • Comparison tables: Use color/weight to highlight surprising findings
  • Interactive visualizations: If feasible, use D3.js or similar (e.g., embedding projections)
  • Conceptual illustrations: Draw analogies visually (e.g., scaling laws as economic investment curves)

Text Hierarchy & Visual Flow

  • Use bold for key concepts (not all caps or colored text)
  • Use italics for variables (\(x\), \(\theta\)) and emphasis
  • Break long text blocks with subheadings, bullet lists, or visual separators
  • Ensure sufficient whitespace (no dense walls of text)
  • Number equations if they’re referenced; use a clear caption

RIGOR & REFERENCES: Academic Integrity

Citations & Attribution

  • Cite primary sources (go to the original paper, not the summary)
  • Cite liberally: Even for well-known concepts, link to the foundational work
  • Include inline links to papers (via DOI or arXiv) where possible
  • For each major claim, ensure a citation or explicit caveat (“To our knowledge…”)
  • Acknowledge related blog posts and prior art; don’t reinvent wheels

References Section

  • Include a full References section at the end with proper formatting
  • Provide BibTeX snippets for academic readers (formatted code block)
  • Organize by theme or chronology (your choice, but be consistent)
  • Verify all links work 1 week before publishing

Reproducibility & Examples

  • If presenting empirical results, link to code or datasets (GitHub, Hugging Face)
  • For algorithms, include pseudocode or reference implementation links
  • Note computational requirements (GPU hours, memory, etc.)
  • If results depend on hyperparameter tuning, be transparent about search methodology

ENGAGEMENT: Build a Conversation, Not a Lecture

Opinionated Positioning

  • Take a stance: What is an unconventional view you’re defending?
  • Address counterarguments directly: “One might think X, but here’s why Y…”
  • Separate facts from interpretation: Use phrases like “We interpret this as…” or “This suggests…”
  • Invite critique: “Does this model hold up in your domain? Let us know in the comments”

Accessibility Without Dumbing Down

  • Signal the audience level upfront: “This assumes familiarity with transformers” or “We explain each step”
  • Provide sidebars or footnotes for tangential depth (keep main narrative uncluttered)
  • Use “show your work” pedagogically—walk through one concrete example fully
  • Recap frequently for long posts: “To recap: X, Y, and Z set the stage for…”

Emotional & Intellectual Resonance

  • Why does this matter beyond the paper? Connect to broader implications
  • Acknowledge limitations honestly: Strengthens credibility
  • Thank collaborators or influences: “This work builds on conversations with…”
  • Use occasional humor (dry, self-aware, not forced)
  • End with a question or open problem to seed follow-up thinking

BEFORE PUBLISHING: The Final Pass

Clarity & Flow

  • Read aloud to catch awkward phrasing, run-on sentences, or unclear jargon
  • Share with a peer unfamiliar with the topic and get feedback on clarity
  • Check narrative coherence: Does each section follow logically from the previous?
  • Verify consistency: One term per concept; consistent notation; tone alignment

Technical Accuracy

  • Verify all equations: Check derivations; ensure notation is consistent
  • Test code examples: Run all snippets; verify outputs
  • Double-check empirical claims: Are numbers/citations correct?
  • Peer review if possible: Have an expert in the domain review technical sections

Formatting & Accessibility

  • Mobile testing: Check layouts, equations, code blocks on small screens
  • Verify all links: Test internal anchors and external URLs
  • Check math rendering: Equations should display cleanly (KaTeX or MathJax)
  • Image quality: All figures at ≥150 DPI; readable text labels

SEO & Discoverability

  • Descriptive title: Include key concepts (e.g., “Scaling Laws Beyond Chinchilla: LLM Efficiency…” not “Thoughts on Scaling”)
  • Summary paragraph: First 2–3 sentences capture the gist for social sharing
  • Add header-level keywords naturally (don’t keyword-stuff; it’s obvious)
  • Use descriptive alt text for images

Final Polish

  • Proofread for grammar and typos (use a tool like Grammarly or Hemingway Editor)
  • Ensure consistent capitalization and punctuation
  • Remove any placeholder text or draft comments
  • Validate YAML front matter (date, tags, author)

POST-PUBLISHING: Maintenance & Feedback

Versioning & Updates

  • Add publication date prominently (at the top or in metadata)
  • If updating: Add a dated “Update” note explaining what changed and when
  • Link to subsequent related posts as they’re published
  • Periodically audit references: Fix broken links within 3 months of publication

Engagement & Iteration

  • Monitor feedback and discussions
  • Correct errors transparently: “We corrected X on [date]”
  • Note if new papers supersede claims: “Recent work (2026) has shown…”
  • Celebrate thoughtful engagement: Reply to substantive comments

Meta: How This Post Should Feel

To a practitioner: “Concrete, useful, not overwhelming. I can apply or test the ideas by end.”

To a researcher: “Rigorous, novel perspective. Connects to my work. Honestly acknowledges limitations.”

To an ICLR organizer: “Accessible yet technical. Advances thinking. Well-executed. Invites discussion.”

To an AI/ML generalist: “I don’t need a PhD to get value, but there’s depth for those who dig.”