research-paper-writing

name	research-paper-writing
description	ML paper pipeline: experiment design to submission.
allowed-tools	["bash","write_file","read_file","web_search","browse_page"]
enabled	true
related-skills	["arxiv","academic-paper-review","plan"]
license	MIT
author	Adapted from hermes-agent (Nous Research, MIT); Orchestra Research

Research Paper Writing Pipeline

Overview

End-to-end pipeline for producing publication-ready ML/AI research papers targeting NeurIPS, ICML, ICLR, ACL, AAAI, COLM. Covers the full research lifecycle: experiment design, execution, analysis, paper writing, review, revision, and submission.

This is not a linear pipeline — it is an iterative loop. Results trigger new experiments. Reviews trigger new analysis.

When to Use

• User is writing an ML/AI research paper for a top venue
• User needs help with experiment design, execution, or analysis
• User wants feedback on a draft
• User is preparing a submission package

Pipeline Phases

Phase 0: Project Setup → Phase 1: Literature Review
       │                        │
       ▼                        ▼
Phase 2: Experiment      Phase 5: Paper Drafting ◄──┐
       Design                  │                    │
       │                       ▼                    │
       ▼                 Phase 6: Self-Review       │
Phase 3: Execution            & Revision ───────────┘
       & Monitoring              │
       │                         ▼
       ▼                   Phase 7: Submission
Phase 4: Analysis

Phase 0: Project Setup

• Define research question and hypothesis
• Identify target venue + deadline
• Set up project structure:

  project/
  ├── experiments/
  ├── data/
  ├── src/
  ├── paper/
  │   ├── main.tex
  │   ├── figures/
  │   └── references.bib
  └── README.md
  

• Initialize git repo, set up environment

Phase 1: Literature Review

• Use arxiv skill to find related work
• Use web_search for non-arXiv papers (Semantic Scholar, Google Scholar)
• Use browse_page to read key papers in full
• Build a references.bib with all cited works
• Identify the gap your work fills

Phase 2: Experiment Design

• Define baselines and comparison methods
• Choose datasets and evaluation metrics
• Design ablation studies
• Plan computational budget
• Write pre-registration document (optional but recommended)

Phase 3: Execution & Monitoring

# Run experiments
python src/train.py --config configs/exp1.yaml

# Monitor with logging
python src/train.py --config configs/exp1.yaml --log-dir runs/exp1

# Track experiments
python src/eval.py --checkpoint runs/exp1/best.pt --eval-set test

• Log all hyperparameters, seeds, and environment details
• Save checkpoints for reproducibility
• Run each experiment with multiple seeds (3-5)

Phase 4: Analysis

• Aggregate results across seeds
• Compute statistical significance (paired t-test, bootstrap CI)
• Generate comparison tables and plots:

  python src/plot.py --results runs/ --output paper/figures/
  

• Run ablation analysis
• Identify surprising findings (investigate, don't hide)

Phase 5: Paper Drafting

Follow venue template structure:

1. Abstract — problem, method, key result, impact (write last)
2. Introduction — motivation, contribution summary, roadmap
3. Related Work — position within literature (from Phase 1)
4. Method — approach, architecture, training procedure
5. Experiments — setup, main results, ablations, analysis
6. Conclusion — summary, limitations, future work

Writing principles:

• One idea per paragraph
• Figures tell the story — design figures first, write text around them
• Tables for comparisons — main results table + ablation table
• Reproducibility — include all hyperparameters, release code

Phase 6: Self-Review & Revision

Use academic-paper-review skill to self-review:

• Read the paper cold (fresh eyes)
• Check methodology soundness, novelty, reproducibility
• Identify weaknesses and fix them
• Get feedback from collaborators

Phase 7: Submission

• Check venue formatting requirements
• Verify page limits
• Anonymize for blind review (if applicable)
• Prepare supplementary material (code, data, extended results)
• Submit before deadline (not at 23:59)

Statistical Analysis

# Multiple seeds — compute mean ± std
python3 -c "
import numpy as np
results = [0.85, 0.83, 0.86, 0.84, 0.82]  # per-seed results
print(f'Mean: {np.mean(results):.4f} ± {np.std(results):.4f}')
"

# Paired t-test vs baseline
python3 -c "
from scipy import stats
baseline = [0.80, 0.79, 0.81, 0.78, 0.80]
ours = [0.85, 0.83, 0.86, 0.84, 0.82]
t, p = stats.ttest_rel(ours, baseline)
print(f't={t:.3f}, p={p:.4f}')
"

Pitfalls

• Single seed: results from one seed are not reliable. Use 3-5 minimum.
• Cherry-picking: report all results, not just the best seed.
• No ablations: reviewers will ask "does each component matter?" — answer proactively.
• Missing related work: reviewers know the field. Cite comprehensively.
• Unclear contributions: list contributions explicitly in the introduction.
• Overclaiming: "state-of-the-art" needs evidence across datasets, not one.
• Last-minute submission: servers crash at deadlines. Submit early.

Dependencies

This skill benefits from: numpy, scipy, matplotlib (analysis + plots). Install via pip install numpy scipy matplotlib.