Algo Reconstruct

Algorithm Reconstructor

Reconstruct methodological logic from papers and code into publication-grade LaTeX pseudocode.

Workflow

1. Gather sources — Read paper methodology section, appendix, and source code. Prioritize paper over code.
2. Identify novelty — Pinpoint the core contribution. Do NOT reduce it to a generic training loop.
3. Select structure — Single unified algorithm, or multi-algorithm by stage. Adapt to paper's natural decomposition.
4. Draft pseudocode — Follow style rules below. Use references/notation.md for symbol mapping.
5. Compile — Run scripts/compile_algo.py on the generated .tex. Fix errors until PDF succeeds.
6. Output — Deliver: Rationale → LaTeX → Complexity Analysis → Compilation Result.

Source Priority

If code conflicts with paper → trust paper, note discrepancy in Rationale.

Abstraction Level

METHOD-LEVEL pseudocode. Include:

• Core methodological pipeline
• Key optimization logic
• Important mathematical transformations
• Novel modules
• Training / inference stages

Exclude:

• Low-level tensor ops (tensor.cuda(), batch.to(device))
• Framework APIs (optimizer.zero_grad(), loss.backward())
• Engineering clutter

If the contribution IS a specific module → elaborate that module sufficiently.

Pseudocode Style

Package: \usepackage[ruled,vlined,linesnumbered]{algorithm2e}

Conventions:

• \KwIn / \KwOut for input/output
• Line numbering enabled
• Scientific notation over programming notation (see references/notation.md)
• Concise mathematical wording

Output Format

Always output in this order:

# Rationale

• How algorithm was reconstructed
• Abstraction decisions and assumptions
• Paper-code discrepancies

# LaTeX Pseudocode

Full compilable LaTeX snippet. Must compile directly.

# Complexity Analysis

• Algorithm intuition (2-3 sentences)
• Time complexity: O(...)
• Space complexity: O(...)
• State assumptions if derivation is uncertain

# Compilation Result

• Engine used
• Success/failure
• PDF path
• Fixes applied

Quality Checklist

Before finalizing, verify:

• Looks like top-tier conference pseudocode
• Reflects methodology, not source code
• Highlights novelty
• Scientifically concise
• Mathematical notation normalized
• No engineering clutter
• Fully compilable
• Publication-ready

Resources

• references/notation.md — Symbol mapping and LaTeX conventions
• references/style_guide.md — Per-venue style expectations and examples
• scripts/compile_algo.py — Compile .tex to PDF with auto-retry