research-paper-writing

Guides end-to-end ML research paper production from experiment design through conference submission.

2|Updated Oct 20, 2017
One-click install
npx skills add https://github.com/rbudiharso/dotfiles --skill research-paper-writing-rbudiharso
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: research-paper-writing
Source: https://github.com/rbudiharso/dotfiles/tree/main/hermes/.hermes/skills/research/research-paper-writing
Command: npx skills add https://github.com/rbudiharso/dotfiles --skill research-paper-writing-rbudiharso

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires semanticscholar, arxiv, habanero, requests, scipy, numpy, matplotlib, SciencePlots, and includes references (resource) and assets (resource) components.

What problem does it solve? Writing a publication-ready ML/AI paper involves coordinating experiments, verified citations, statistical analysis, LaTeX formatting, and venue-specific checklists, and mistakes like hallucinated citations or missing NeurIPS checklist items cause desk rejections. ## Core Features & Use Cases - Full Research Lifecycle: Covers project setup, literature review, experiment design, execution monitoring, statistical analysis, drafting, self-review, and submission for NeurIPS, ICML, ICLR, ACL, AAAI, and COLM. - Verified Citation Workflow: Fetches BibTeX programmatically via Semantic Scholar, CrossRef, and arXiv APIs instead of generating citations from memory, with a mandatory 5-step verification process. - Venue Templates & Checklists: Ships official LaTeX templates for six conferences plus pre-submission checklists covering page limits, anonymization, limitations sections, and LLM disclosure rules. - Use Case: A researcher with experimental results asks the agent to draft an ICML submission; the skill produces a grounded draft using the experiment log, verified references, the icml2026 template, and statistical significance tests. ## Quick Start Ask the agent to help write a research paper from your existing experiment results, specifying the target venue such as NeurIPS or ICML.

Frequently Asked Questions about research-paper-writing

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I write a machine learning paper for NeurIPS or ICML?▼

Follow a phased pipeline: define a one-sentence contribution, map every experiment to a claim, run baselines with statistical significance tests, then draft using the official venue LaTeX template. Complete the venue's mandatory checklist before submission to avoid desk rejection.

How do I prevent hallucinated citations in AI-assisted paper writing?▼

Never generate BibTeX from memory; fetch it programmatically via DOI content negotiation from CrossRef or the Semantic Scholar API. Verify each paper exists in at least two sources and confirm the cited claim appears in the abstract, marking unverifiable entries as placeholders.

What LaTeX templates are included for conference submissions?▼

The templates directory includes official style files for ICML 2026, ICLR 2026, NeurIPS 2025, ACL, AAAI 2026, and COLM 2025, each with the .sty file, bibliography style, and example document. Compile with pdflatex plus bibtex or latexmk.

Does this workflow support human evaluation studies for NLP papers?▼

Yes, it covers annotator selection, pairwise versus Likert scale design, inter-annotator agreement metrics like Krippendorff's alpha, crowdsourcing platform choice, and compensation reporting. It also includes IRB guidance and the reporting details ACL and EMNLP reviewers expect.

What statistical tests should I report for ML experiment comparisons?▼

Use McNemar's test for paired binary outcomes between two methods, bootstrapped 95% confidence intervals for key metrics, and Cohen's h for effect size. Always report sample sizes, number of runs, and whether error bars show standard deviation or standard error.

When should I use autoreason iterative refinement versus a single pass?▼

Use autoreason for mid-tier models on constrained tasks where the generation-evaluation gap is large; it wins consistently there. For frontier models on unconstrained tasks or very weak models, single pass or critique-and-revise performs better and costs far less compute.