Research Paper Writing

• Starting a new research paper from an existing codebase or idea, including setting up workspace structure, version control, and a one-sentence contribution statement before writing anything.
• Designing and running experiments that map directly to paper claims, with incremental saving, artifact preservation, cost tracking, and monitoring via cron-scheduled status checks.
• Writing or revising any section of an ML/AI paper, including generating LaTeX using official conference templates (NeurIPS, ICML, ICLR, ACL, AAAI, COLM) with a two-pass refinement pattern.
• Preparing for submission or resubmission, including anonymization, formatting verification, pre-compilation validation, conference-specific checklists, and post-acceptance deliverables such as posters, talks, and code releases.

• Non-ML/AI research domains (e.g., medicine, social sciences, humanities) — the pipeline is specifically calibrated for ML conference venues and their norms.
• Very weak base models (Llama 8B class) for iterative refinement tasks — the skill notes these models are too weak for diverse candidate generation, making the autoreason loop counterproductive.
• Quick one-off writing tasks with no experimental component — the overhead of the full 8-phase pipeline outweighs the benefit for small, self-contained writing jobs.

Covers Project Setup → Literature Review → Experiment Design → Execution & Monitoring → Analysis → Paper Drafting → Self-Review & Revision → Submission Preparation as an iterative loop, not a linear sequence. Results trigger new experiments; reviews trigger new analysis; the skill handles these feedback loops explicitly.

Every citation is fetched programmatically via Semantic Scholar, arXiv, and CrossRef DOI content negotiation—never generated from memory. Unverifiable citations are marked [CITATION NEEDED] and reported to the scientist, addressing the known ~40% AI citation error rate.

An evidence-based iterative refinement loop (Critic → Author B → Synthesizer → 3-judge Borda panel) with empirically validated parameters: k=2 convergence, CoT judges, temperature 0.8 authors / 0.3 judges. Includes a decision table mapping model tier and task type to the optimal refinement strategy, plus documented failure modes and mitigations.

Generates N=3–5 independent reviews with a negative-bias prompt, then aggregates via a meta-reviewer role modeling an Area Chair. Includes a separate VLM-based visual review pass for figure quality and layout issues, and a claim verification pass using a fresh sub-agent to prevent confirmation bias.

Includes ready-to-use templates for NeurIPS 2025, ICML 2026, ICLR 2026, ACL, AAAI 2026, and COLM 2025, with a professional preamble (microtype, booktabs, siunitx, cleveref, algorithm2e, TikZ, SciencePlots), pre-compilation validation scripts (chktex, citation/figure/label checks), and a latexdiff workflow for rebuttal revision tracking.

Designed for the Hermes agent: uses delegate_task for parallel section drafting and concurrent citation verification, cronjob for experiment monitoring with a [SILENT] protocol to suppress no-change notifications, memory and todo for persistent state across sessions, and send_message for async notifications when experiments complete.

Starting from an existing repository, the skill explores the codebase to identify the contribution, designs experiments that map to specific claims, runs them with incremental checkpointing, analyzes results with statistical significance tests, drafts a complete LaTeX paper using the target venue's template, and prepares the final anonymized submission package.

After a rejection, the skill converts the paper to a new venue's format (including page-limit adjustments and venue-specific required sections), addresses reviewer concerns in the revised text, generates a latexdiff marked-up PDF showing changes, and verifies the new submission against the target venue's checklist—without referencing the previous submission.

For papers requiring human evaluation as primary evidence (e.g., ACL generation tasks), the skill designs the annotation protocol—annotator type, scale (pairwise vs. Likert), sample size via power analysis, inter-annotator agreement metric selection, platform choice (Prolific, MTurk), and IRB checklist—before running automated experiments, since human eval typically has longer lead times.

Beyond empirical ML, the skill supports theory papers (theorem/proof structure with proof sketches in main text and full proofs in appendix), survey papers (breadth-first literature search with taxonomy design), benchmark papers (dataset documentation via Datasheets for Datasets, construct validity evidence), and position papers—each with distinct structure and evidence standards.