biswajit56546/academic-slide-refiner

Academic Slide Refiner

This skill provides a structured workflow for transforming a dense, paper-like presentation into a clear, concise, and compelling talk for a short academic conference slot (e.g., 20 minutes).

Core Principles

1. One Core Idea Per Slide: Every slide must have a single, clear takeaway. If you have two ideas, make two slides.
2. The 5-Second Rule: An audience member, seeing the slide for the first time, should grasp its main point within 5 seconds without needing to read the whole paper.
3. Talk, Don't Read: Slides are visual aids for the audience, not speaker notes. Use minimal text. The speaker provides the detail.
4. Strict Information Diet: Adhere to these limits:
   • Maximum 3–4 bullet points per slide.
   • Maximum 2 lines of text per bullet point.
   • No long, complex, paper-style sentences.
5. Visual Consistency: A unified style is critical for a professional look. Stick to a simple, high-contrast theme.
   • Default Style: White background, dark text (black or very dark blue), and one or two accent colors for highlighting.
   • No gradients, shadows, or decorative backgrounds.

External Report Constraints

For advisor-facing decks:

1. Use English only.
2. Remove internal/local markers from final slides:
   • no absolute paths or local directory names,
   • no long version tags (*_20260304_192047 style),
   • no internal host/user identifiers.
3. Prefer semantic experiment labels (Baseline, Variant-B, Targeted-BM) over raw run names.
4. If output format is requested as PDF, prioritize beamer (.tex -> .pdf) and keep a source .tex artifact.

Refinement Workflow

Re-organize the entire presentation to follow a clear, logical narrative. A highly effective structure for technical talks is:

Problem → Challenges → Key Mechanism → Why It Works → Experimental Validation → Conclusion

Ensure there are clear transition statements or pages connecting each stage of this narrative.

Step 1: Restructure and Split

• Re-order slides to fit the causal chain above.
• Split complex pages. A single dense slide should be broken down into multiple simpler slides. For example:
  • A slide explaining VQ-VAE should be split into: 1. What is VAE? → 2. What is VQ-VAE (Discrete Latent)? → 3. Why VQ-VAE Helps Our Problem?
  • A slide on Bayesian Optimization should be split into: 1. Why BO is Needed (e.g., Non-differentiable) → 2. How BO Works in Our Context (e.g., in Latent Space).

Step 2: Content & Layout Refinement

• Text: Ensure font size is large enough for back-of-the-room visibility (>= 22pt is a good rule of thumb).
• Layouts: For common "left text, right image" layouts, ensure content is vertically centered to create visual balance. Avoid having a small image next to a large block of text, making the page look lopsided.
• Images & Diagrams:
  • Images should be large and clear, occupying 40-60% of the slide area.
  • Ensure diagrams (e.g., model architectures) are stylistically consistent in terms of color, font, and line weight. Use the bundled scripts as a starting point for VAE/VQ-VAE diagrams.
  • Verify all image paths are correct to prevent broken images showing up as text.
  • All figures must have clear, numbered captions (e.g., "Fig. 1(a): ...").

Step 3: Enforce Academic Rigor

• Define Acronyms: Spell out all acronyms on their first appearance. For example: MMM (Machine Learning-Based Macro-Modeling).
• Add References: For all prior work mentioned (e.g., ALIGN, MMM), add a citation number [X] in the text. At the bottom of the slide, add a corresponding reference footnote.
• Re-number References: The reference numbering should follow the order of appearance in the slides, not the order from the original paper.

Page-Specific Best Practices

• Motivation Pages: Clearly show the problem. Use contrasting visuals (e.g., Schematic vs. Layout performance) to highlight the performance gap. Explicitly state why existing solutions (e.g., HPWL) are insufficient.
• "Why X in Y?" Pages: The focus should be on the synergy. The question is not "Why use Bayesian Optimization?" but "Why is Bayesian Optimization effective in Latent Space?". The key is to explain the advantages gained from the combination (e.g., Feasibility Guarantee, Dimensionality Reduction).
• Limitations Pages: When discussing the limitations of prior work (e.g., MMM), be direct and use simple language. A text-only, centered layout with large fonts can be very effective for impact.

Bundled Resources

• scripts/gen_vae_diagram.py: A Python script using Matplotlib to generate a VAE architecture diagram with a clean, academic style.
• scripts/gen_vqvae_diagram.py: A Python script to generate a VQ-VAE architecture diagram, stylistically matching the VAE diagram.