Write a research abstract as LaTeX, with citations
In this exercise, you direct the agent to synthesize empirical results and verified bibliographic citations into a publication-ready structured scientific abstract in LaTeX (out/abstract.tex), compile the document to PDF, and trace every numerical claim back to underlying data tables.
Before you start: A results file from statistical analysis, a verified references.bib containing retrieved citations, and a local LaTeX compiler environment (Setup). If LaTeX is absent, the agent can configure it (Environment), or compile via Markdown and Pandoc.
Background
Adopting a source-first compilation pattern ensures maximum reproducibility (Outputs). LaTeX generates clean, standardized typography with automated section numbering and bibliographic citation formatting resolved against references.bib. Because the final PDF is compiled programmatically from plain text, updating estimates or refining prose requires only a concise source edit and recompilation.
Two mandatory guardrails maintain scientific integrity:
- Numerical claims derive strictly from results files: Every percentage, odds ratio, and sample size must be extracted directly from
out/results.mdand flow tables, rather than recalled from conversational memory. - Citations derive strictly from
references.bib: The agent is prohibited from introducing external citations that do not exist within the local bibliography (Verification).
A structured scientific abstract adheres to standard sections: Background, Methods, Results, and Conclusions. Ensure the concluding interpretation aligns strictly with what the observational design mathematically supports.
General methodology
Direct the agent to draft the structured abstract, embed citations from the verified bibliography, compile the document, and provide an explicit provenance checklist for every number:
Draft a single-page structured abstract (Background, Methods, Results, Conclusions) as out/abstract.tex. Incorporate an embedded results summary table and citations drawn exclusively from references.bib. Extract all numerical estimates, denominators, and confidence intervals directly from <results file> and <flow table>. Do not fabricate or introduce any citation key absent from references.bib. State the study design clearly, noting primary observational limitations. Compile the document to out/abstract.pdf and report any LaTeX compiler warnings. Finally, provide a provenance checklist mapping every number and citation key in the abstract to its specific source table or reference record.
Workshop example · Imaging: chest radiographs
Start from: the referenced checkpoint (results plus references.bib) (how to restore)
Produces: out/abstract.tex, out/abstract.pdf
Prompt
Write a one-page structured abstract (Background, Methods, Results, Conclusions) as out/abstract.tex, for the department's AI committee, about whether the open-source pneumothorax detectors are ready for a pilot. Take every number from out/evaluation.md and out/cohort_flow.md, and the cleaning counts from out/cleaning_flow.md. Cite only entries in references.bib, and do not add any other citation. Include one results table, state that this is a retrospective evaluation on a small public sample with synthetic data problems and that the models may have seen these images in training, and give the main limitation. Compile it to out/abstract.pdf with the bibliography and report any warnings. Then list each number and each citation with its source.
Files used: out/evaluation.md, out/cohort_flow.md, out/cleaning_flow.md, references.bib.
Reference result. The reference abstract is scenarios/imaging/checkpoints/reported/out/abstract.pdf, one page. Its key numbers:
| Item | Value |
|---|---|
| Images received / cleaned | 283 / 258 |
| Cohort | 203 adult patients, 59 with pneumothorax (PA 120, AP 83) |
AUC of the primary model (all) |
0.78 (95% CI 0.72 to 0.84) |
| Other models | chex 0.72, nih 0.67, pc 0.69 |
Primary minus nih / pc / chex |
+0.11 (0.04 to 0.18) / +0.09 (0.01 to 0.16) / +0.06 (-0.02 to 0.15) |
| Primary at its published threshold | sensitivity 0.92, specificity 0.47 |
| Pilot bar (sens at least 0.85, spec at least 0.80) | Not met by any model |
| PA vs AP (exploratory) | AUC 0.85 vs 0.74, difference 0.11 (-0.03 to 0.26) |
Compare yours. Your wording will differ. Check the numbers against this table and the citations against references.bib. The conclusions should not claim more than the table does: no model met the bar, and the PA vs AP difference is not established.
Check the citations by reading. Ask: "For each citation, quote the sentence in the paper's abstract that supports the claim." If the agent cannot, the citation is decoration.
Verification criteria
- Numerical concordance: Every statistic, sample size, and confidence interval in the PDF matches underlying results tables exactly.
- Bibliographic validity: All citation keys resolve cleanly against
references.bibwithout missing reference warnings (?or[?]), and cited papers substantiate their corresponding assertions. - Epistemic calibration: The conclusions section discusses observed statistical associations without overreaching into unverified causal claims.
- Synthetic data disclosure: If analyzing the workshop benchmark dataset, the abstract explicitly discloses that the findings derive from synthetic data.
Common pitfalls
| Symptom | Underlying cause | Resolution |
|---|---|---|
Unresolved citations ([?]) |
The bibliography engine did not execute the full compile cycle | Recompile using the standard sequence: pdflatex, bibtex, followed by two passes of pdflatex |
| Missing LaTeX package errors | Required style packages are absent from the host LaTeX distribution | Authorize the agent to install missing packages via the system package manager, or simplify the LaTeX preamble |
| Disconnected citations | The model attached a legitimate citation to an unrelated claim | Direct the agent to cite the supporting sentence from the cited paper's abstract, or remove the reference |
| Document spills onto a second page | Preamble geometry or margins exceed single-page limits | Instruct the agent to adjust geometry margins or tighten section spacing |
Concepts behind this
- Outputs. Request text-based source files and compile them to presentation formats. Source files are inspectable, version-controlled, and reproducible. Go deeper →
- Verification. Language models can be confidently incorrect. Require empirical evidence, verify mathematical invariants, and use clean-context peer review to audit findings. Go deeper →
Builds on: Evaluate an AI tool, with intervals