| name | academic-paper-writing |
| description | Guide and draft academic papers (high-energy physics and machine learning focus) following a rigorous section-by-section structure: Abstract, Introduction, Method, Results, Discussion, Conclusion. Use when the user needs to write, review, restructure, or improve an academic manuscript — including drafting sections, critiquing drafts against the structure below, or generating outline scaffolds. |
Academic Paper Writing
Overview
This skill provides a prescriptive structure for writing research papers, primarily targeting high-energy physics (HEP) and machine learning (ML) audiences. Each section has a defined internal logic. Follow it both when drafting from scratch and when critiquing an existing draft.
Paper Structure
1. Abstract
The abstract must cover five elements in order. Do not merge or reorder them.
| Part | Guidance |
|---|
| a. Background | State the current landscape and its deficiency. Two components: (1) what is known / what exists; (2) what is missing, insufficient, or problematic. |
| b. Purpose | Directly respond to the deficiency stated in (a). The purpose sentence must be traceable to the gap — do not introduce new motivation here. |
| c. Method | Concisely name the approach, dataset, and key design choices. |
| d. Results | State the quantitative or qualitative outcome. Be specific — include numbers when the journal/venue permits. |
| e. Conclusion | Abstract and elevate the results to a higher-level insight. Do not repeat the result sentence verbatim. The conclusion should answer: "What does this finding mean for the field, the theory, or future work?" |
2. Introduction
The introduction follows a five-step funnel.
a. Research background and purpose
Open with the broad scientific or technical context. State the research objective early so the reader knows the destination before the literature review begins.
b. Related work, expected outcomes, and current status
Identify the specific sub-area(s) relevant to the objective. Describe what has been attempted, what is expected from theory or prior experiment, and where the field currently stands.
c. Systematic literature review
Build on (a) and (b) to survey the literature in relation to the stated objective. Requirements:
- Organize references into a logical taxonomy: by research sub-field, by methodological approach, by chronological development, or by any other principled scheme — but never as a flat sequential list.
- Synthesize, do not annotate. For each group, summarize the collective finding and its significance, not a paper-by-paper description.
- Use the literature to establish what is established, what is contested, and what is merely assumed.
d. Research gap
Based on (c), identify the specific gap: what question the community has not yet answered, what limitation no existing method addresses, or what contradiction remains unresolved. The gap must follow logically from the review — it should feel inevitable, not asserted.
e. This paper's contribution
State concisely: (1) which gap this paper addresses, (2) the high-level research design (study type, methodology, data), and (3) the scope. End with a brief roadmap of sections if the journal convention requires it.
3. Method
Follow the conventions of the relevant sub-field:
High-Energy Physics (HEP):
- Detector description and data-taking conditions (beam energy, luminosity, run period)
- Monte Carlo simulation and generator settings
- Event selection and cut flow with efficiency and background rejection
- Observable definitions, kinematic variables, and units
- Systematic uncertainty sources and estimation procedure
- Statistical analysis framework (likelihood, frequentist / Bayesian choice)
Machine Learning:
- Dataset description (source, size, splits, preprocessing, class balance)
- Model architecture and key design choices with justification
- Loss function and training objective
- Optimization: optimizer, learning rate schedule, regularization, early stopping
- Evaluation metrics and their definition
- Baseline models and comparison protocol
- Implementation details: framework, hardware, training time, random seeds
For combined HEP+ML papers (e.g., particle reconstruction with deep learning), cover both blocks in logical order — detector and data first, then the ML pipeline applied to that data.
4. Results
Figures and tables
- Every figure and table must be self-contained: a reader who skips the body text must understand what is shown from the caption alone.
- Caption requirements: (1) a specific descriptive title — not "Figure 1" but "Energy resolution as a function of pseudorapidity for the upgraded calorimeter"; (2) all symbols, colors, and line styles defined; (3) units stated; (4) any selection or condition applied to produce the plot noted.
- Axis labels must include units. Legends must not overlap data.
Describing results
Use a topic-sentence-first (general → specific) structure for each result block:
- Topic sentence: one sentence stating the main finding of this result block.
- Supporting description: walk through the evidence — panel by panel, row by row, or condition by condition — to substantiate the claim.
- Quantitative anchors: every non-trivial claim should cite a number (value ± uncertainty, percentage improvement, significance in σ, etc.).
Do not interpret or explain results in this section. Save causal arguments and comparisons to prior work for Discussion.
5. Discussion
The discussion has four required paragraphs (or paragraph groups). Do not collapse them.
a. Summary of principal findings (paragraph 1)
Open with a concise restatement of the main results. This paragraph serves readers who skip from Abstract directly to Discussion. Keep it tight — two to four sentences.
b. Comparison with prior work (paragraph 2)
For each principal finding:
- Find relevant published results and compare directly (same observable, similar conditions).
- If results agree: cite and briefly explain why agreement is expected.
- If results disagree: explicitly state the discrepancy, then analyze the likely causes (different dataset, different selection, different model architecture, different systematic treatment, different definition of the observable, etc.). Do not downplay disagreements.
c. Practical significance and generalizability (paragraph 3)
Elevate the findings beyond the immediate experiment:
- What does this mean for practitioners, experimentalists, or theorists?
- How broadly does the result generalize? (Different detectors, different energy scales, different datasets, different domains?)
- What design decisions or policy choices does this inform?
d. Limitations (paragraph 4)
State the boundaries of the result honestly:
- Data limitations (statistics, coverage, selection bias)
- Model or method limitations (assumptions, approximations, known failure modes)
- What the study does not address that a reader might expect
- What follow-up work is required to extend the validity
Do not frame limitations as future work. Acknowledge them as current constraints.
6. Conclusion
- Synthesize the overall contribution in three to five sentences.
- Do not introduce new results or new citations.
- End with the broader implication or the most important open question the work raises.
- The conclusion is not a second abstract — it should read as a forward-looking synthesis, not a backward-looking summary.
Cross-Cutting Rules
Language and style
- Prefer active voice for method descriptions; passive voice is acceptable for results when the subject is the observable rather than the authors.
- Avoid hedging chains: "it may be possible that" → "this suggests".
- State uncertainty explicitly (statistical vs. systematic; aleatory vs. epistemic) rather than using vague qualifiers.
Citation discipline
- Cite the original source, not a review that cites it.
- When a claim is common knowledge in the sub-field, no citation is needed; when it is not, cite it.
- Do not cite papers you have not read. If you are uncertain about a reference, flag it for the user to verify.
Consistency checks before submission
- Abstract conclusion ↔ Discussion paragraph (c): same level of claim?
- Introduction gap ↔ Contribution statement: does the paper address exactly the gap it identified?
- Method ↔ Results: every reported metric must be defined in Method.
- Results numbers ↔ Abstract numbers: must match exactly.
- All figures cited in text in the order they appear.
