Manifold Grace
RESEARCH & KNOWLEDGE · RESEARCH SKILLS

Doctoral AI Research Scientist: Skills & Paper Standards

A practical, discipline-sensitive standard for developing rigorous AI research papers, theoretical contributions, reproducible software, empirical studies and responsible scientific judgement.

RESEARCH SKILLS REFERENCE · 9 NOVEMBER 2026

Purpose and scope

These competencies reflect common expectations in advanced AI research and widely recognised publication practices. They are a developmental reference, not a doctoral qualification, a university-approved syllabus or a claim that every paper requires all techniques. Research excellence is demonstrated through methods suited to the question, credible evidence, transparency and a defensible original contribution.

Twelve skills for doctoral-level AI research

CapabilityApplied research skillEvidence of competent practice
01 · Problem formation & research noveltyFormulate falsifiable questions; bound claims; analyse prior work, competing explanations and research contribution.Produce a scoped problem statement, novelty comparison and a hypothesis-to-evidence map.
02 · Literature review & bibliographic integritySearch systematically, follow citation chains, assess source quality, and distinguish preprints from peer-reviewed work.Reproducible search log, comparison matrix, verified DOIs and reference manager exports.
03 · Mathematics & theoretical foundationsLinear algebra, probability, statistical inference, optimisation, information theory, formal logic and mathematical proofs as relevant.Derivations, stated assumptions, proofs or counterexamples; numerical checks.
04 · Computer science & implementationPython, software design, data structures, algorithms, PyTorch/JAX where relevant, version control and automated testing.Reviewed repository, environment file, unit tests, meaningful code comments and repeatable runs.
05 · Experimental and causal designBaselines, controls, counterfactuals, ablations, sampling, leakage checks, preregistration where suitable.Study protocol, baselines, sampling justification and experiments isolating individual effects.
06 · Datasets and measurementCollection provenance, cleaning, inclusion/exclusion criteria, licences, annotation guidelines, uncertainty and representativeness.Dataset description, consent/licence record, quality checks, split design and data dictionary.
07 · Evaluation and statistical reportingAppropriate metrics, confidence intervals, calibration, robustness, error analysis, effect sizes, multiple-comparison awareness.Statistical reports, breakdowns, uncertainty, failure cases and reproducible plots.
08 · Interpretability and epistemic limitsSeparate predictions, observations, causal hypotheses and explanations; identify unknowns and coverage limitations.Provenance graph, missingness register, transparent error taxonomy and interpretability tests.
09 · Safety, ethics and governancePrivacy, fairness, accessibility, legal constraints, security, possible misuse, environmental and societal implications.Risk assessment, ethics approvals where needed, safeguards and limitations.
10 · Reproducibility & research operationsSeed control, software and hardware versions, hyperparameters, experiment tracking, compute and cost accounting.Executable instructions, scripts, run records, model/dataset cards and archival release.
11 · Scientific communication & peer reviewClear abstract; structured introduction; related work; methods; results; discussion; limitations; qualified conclusions.Readable figures/tables, accurate citations, reviewer-response record and journal-specific formatting.
12 · Professional research independenceMaintain integrity, address criticism, protect contributors, collaborate across disciplines and continue specialist learning.Authorship contribution record, peer feedback, versioned corrections and reflective research log.

Publication-quality manuscript architecture

ComponentRequired intellectual work
Title, abstract & keywordsState exactly what is studied, the contribution, approach, evidence and supported findings; do not present planned results as completed.
Introduction & research questionsDefine the problem, significance, gaps and claims with measurable boundaries.
Related workCompare relevant studies, contradictory findings and alternative explanations without overstating novelty.
Theory, model or methodSpecify assumptions, definitions, equations, design decisions, datasets and procedures to the necessary depth.
Experiments / analysesPresent baselines, ablations, evaluation metrics and uncertainty if empirical; complete proofs or arguments if theoretical.
ResultsReport actual outcomes, tables and figures with denominators, units, uncertainty and appropriate comparisons.
Discussion & limitationsAnalyse competing interpretations, generalisability, omitted variables, failure cases and ethical implications.
Conclusion & referencesRestrict conclusions to what evidence supports, present future questions, cite traceable sources and document contributions.
Appendices & artifactsProvide detailed proofs, protocols, code, model and data documentation, and reproduction instructions where appropriate.

Exemplary formatting patterns from the five reference papers

The user-provided example papers illustrate different legitimate types of scholarly presentation. This page abstracts structural lessons; it does not reproduce their contents.

Reference exampleParticularly useful practiceBest suited future papers
Cross-national work-schedules studyPopulation definition, sampling, comparative data analysis and policy interpretation.Social science, education, employment and institutional evaluation.
DeepSeek and LLM comparisonNamed comparison tasks, evaluation protocol, dataset labelling and comparative metrics.LLM benchmarking, evaluation and comparative AI studies.
Origins of representation manifoldsMathematical definitions, explicit hypotheses, propositions and visual diagnostics.Formal theory, ontology and representation learning.
SimpleFoldArchitecture description, equations, benchmark comparisons, figures and linked implementation.Model design, applied machine learning and scientific computing.
rStar2-AgentTechnical contents hierarchy, training stages, experimental and ablation analyses, and unsuccessful attempts.Agentic AI, engineering evaluation and reproducibility reports.

Research paper and coding project relationship

Where a research question can be operationalised, maintain a linked, versioned code project with a README, research question, installation instructions, environment dependencies, permitted dataset sources, tests, reproducible experiment scripts, expected outputs, limitations, model or dataset documentation and explicit licence terms. A GitHub repository is useful for code and version history; a Zenodo record/DOI can archive a specific deposited release. Neither should be cited as an existing artifact before creation or deposition.

For literature reviews and normative work, a transparent selection protocol, evidence register and versioned analysis may be more appropriate than forcing an unnecessary code project.

Quality gates before submission

GateWhat must be checked
Scientific contributionQuestion, prior art, originality claims and value of the contribution.
Technical validityProofs, implementation fidelity, robust baselines, correct experimental design, absence of leakage.
Evidence qualityProvenance, measurement, uncertainty, counterevidence, domain limitations and claims-to-results consistency.
Responsible researchPermissions, confidentiality, privacy, risks, fairness and appropriate institutional review.
ReproducibilityDocumented methods, software and dataset versions, compute details and independent testability.
Publication complianceVenue template, anonymity, pages, references, ethics/reproducibility statements and LLM-use disclosures where required.

A paper may be theoretically rigorous without code, empirically rigorous without inventing a new model, or technically rigorous without an exhaustive social-policy analysis. Match the standards to the contribution and venue.

Current research and industry guidance

NeurIPS Paper Checklist — claims, limitations, theoretical assumptions, reproducibility, ethical and societal impacts. ICLR 2026 Author Guide — venue submission, reproducibility, ethics and disclosure of substantial AI contributions. ACM-style artifact evaluation example — documented, consistent, complete and exercisable research artifacts. NIST AI RMF Playbook — governing, mapping, measuring and managing applicable AI risks. Venue policies differ and should be reviewed at submission time.

Publisher-specific layouts and venue selection: verified reference (November 2026)

Use the official template for the intended venue, rather than making every paper resemble an arXiv preprint. The limits below are version-specific; always recheck the actual submission year's author guide. A venue-fit recommendation is a planning aid, not an indication of acceptance or eligibility.

Venue / formatVerified presentation requirementsMost suitable manuscript familiesOfficial guidance
NeurIPS 2026Official LaTeX; nine content pages plus references, appendices and mandatory checklist.Machine learning methods, AI evaluation and reproducibility.Main-track handbook
ICML 2026LaTeX; eight main-body pages at submission; references and appendices follow; anonymisation required.Algorithms, systems and comparative experimental ML.Author instructions
ICLR 2027Official LaTeX; nine-page main text at submission, ten pages for later stages; double-blind.Representation learning, model architectures and reasoning.Author guidelines
AAAI 2027Two columns, US Letter; up to seven non-reference pages in a nine-page main submission PDF.Knowledge representation, agents and general AI research.Submission instructions
*ACL proceedingsOfficial templates; two-column A4; long-paper review version up to eight content pages.LLM language, retrieval, reliability and evaluation.Official styles · Formatting details
ACM FAccT 2026Interdisciplinary; author guide and approved ACM styles. Camera-ready instructions for 2026 adopted single-column and up to 15 body pages.Algorithmic fairness, public accountability and AI social impact.Author guide · Camera-ready format
JMLRRequired journal LaTeX style, PDF submission, longer methods and technical proofs accommodated.Substantial theoretical, mathematical and empirical ML papers.Author information · Formatting
Nature Machine IntelligenceFlexible initial PDF/Word/LaTeX submission; research Articles typically 3,500 main-text words, 150-word abstract and six figures/tables.Significant, broadly relevant AI discoveries with strong evidence.Article types · Initial formatting
IEEE journalsPublication-specific Word/LaTeX article templates; verify the actual IEEE journal and article category.Applied AI engineering, networks, secure and distributed systems.Authoring tools
PLOS ONEStructured scientific manuscript; results and methods; separately submitted figures and in-text tables with rigorous statistical reporting.Reproducible empirical, interdisciplinary and applied research.Submission guidelines

Recommended adaptable manuscript families

Master manuscriptIntellectual architectureFormatting approach
Theoretical / epistemologicalProblem, prior art, definitions, propositions, counterexamples, implications and limits.Single-column working manuscript, then JMLR, ICLR or other venue template if appropriately scoped.
Experimental AI / benchmarkingHypotheses, datasets, baselines, protocol, metrics, ablations, uncertainty, error analysis and reproducible code.Venue-compliant LaTeX for NeurIPS, ICML, ICLR, AAAI or ACL as topic warrants.
Engineering / systemsRequirements, architecture, implementation, security and failure modes, evaluation, reproducibility and limitations.IEEE or relevant ACM/AI venue template; diagrams and executable artifact record.
Social science / policyLiterature review, conceptual model, sampling, policy chronology, quantitative or qualitative methods, findings and limitations.Discipline-appropriate journal template, or FAccT for rigorous interdisciplinary AI governance.
Critical review / comparative analysisScope, search strategy, eligibility, inclusion decisions, extraction protocol, appraisal, synthesis and evidence gaps.Journal-specific review or analysis format; disclose when the work is a scoping rather than systematic review.
Working paper / preprintTransparent manuscript status, original scope, methods or argument and complete references.Readable independent layout until a venue is selected; archives are dissemination channels, not endorsement or peer review.

Best-practice layout and reproducibility tests

Use the publisher's original font, margins, columns, page size, headings and citation rules without custom restyling. Verify readable vector figures, accessibility of labels without colour alone, complete table units, embedded fonts, equations, bibliography data, anonymised author metadata where required, and aligned text/figure citations. Confirm that abstract, claims and actual findings agree. Ensure that significant comparisons include proper controls, and report missing data, uncertainty, limitations, research ethics, use of AI tools and code/data availability as the applicable venue requires.

Publication-integrity rule: GitHub is for versioned code and research artifacts; Zenodo and preprint servers are for appropriate public archives, not replacements for journal or conference peer review. Do not make a paper publicly visible during double-blind evaluation without checking venue rules; avoid false published/accepted status.

Editorial quality: Nature Machine Intelligence explicitly weighs methodological soundness, support for conclusions, significance and broad relevance, not just document appearance. Editorial process. NeurIPS requires a transparency/reproducibility checklist. Paper checklist.

Download manuscript starter templates

Access six editable research manuscript templates for theory, experiments, engineering, policy, evidence reviews and working-paper initiatives. These are independent starter structures; use the actual publisher template when submitting.

Connected research

Learning Lab · Education, Learning & Development · Research · Research Questions & Gaps · Frameworks Completeness Atlas.

This reference can inform paper-specific skills development, manuscript planning and technical project records without implying that a doctorate or research publication has been awarded.

Research skills reference created 9 November 2026. Standards are linked to their respective issuers; individual submission requirements may change.