Write and Verify ML Research Papers with Hermes

The bundled Research Paper Writing skill is a complete project pipeline, not a request to generate academic-sounding prose. It starts with repository and contribution discovery, proceeds through literature review and experiment design, preserves raw results and costs, and ends with reviewer-style critique and submission preparation.

Official skill reference: Research Paper Writing

Make the contribution testable

Write the one-sentence contribution and the claims each experiment must support. Search related work in multiple rounds: broad terminology discovery, deeper citation tracing, and targeted searches for missing baselines or recent competing work. Verify each citation against live metadata and the actual paper.

Design experiments before running them. Specify datasets or tasks, baselines, metrics, seeds, compute budget, stopping conditions, and result paths. Save each completed unit so interrupted jobs can resume without repeating expensive work. Preserve failed and negative experiments; they constrain the honest story.

Paper workflow checklist

• State one falsifiable contribution in a sentence.
• Build a verified bibliography and related-work map.
• Connect every central claim to an experiment or source.
• Predefine baselines, metrics, seeds, and statistical analysis.
• Log model, data, configuration, cost, and result provenance.
• Generate figures from saved results, not hand-entered values.
• Simulate skeptical reviewers and resolve major objections.
• Rebuild the final PDF and inspect every page.

Common pitfalls

• Writing the conclusion before the evidence exists. Let results shape the claim.
• Choosing weak baselines. Compare against credible current and simple alternatives.
• Reporting only the best run. Preserve seeds, variance, exclusions, and failure criteria.
• Inventing citations or BibTeX. Retrieve and verify metadata programmatically.
• Treating a compiled PDF as a valid submission. Formatting, anonymity, limits, and supplemental rules still need checks.

Verification

Re-run analysis from raw results in a clean environment and compare generated tables and figures with the paper. Check every number, citation, cross-reference, appendix pointer, and claim-to-evidence link. Read the PDF at normal size for clipped figures or tiny labels, then run a final reviewer pass focused on novelty, reproducibility, limitations, and overclaiming.