experiments-writer

Experiments Writer

Purpose

Draft or revise the Experiments section so that the evidence directly answers the paper's claims. Experiments should be a proof plan, not a collection of convenient results.

When To Use

• Planning experiments before submission.
• Revising an experiments section after reviewer-style feedback.
• Checking baseline, metric, dataset, and ablation coverage.
• Preparing a venue-specific empirical story.

Inputs

Required:

• claim-evidence matrix;
• datasets and splits;
• baselines;
• metrics;
• implementation and training protocol;
• main results;
• ablations.

Optional:

• statistical tests;
• seed variance;
• qualitative examples;
• failure cases;
• compute budget.

Procedure

1. Map each main claim to an experiment or analysis.
2. Define datasets, splits, metrics, and evaluation protocol before reporting numbers.
3. Justify baselines and include recent target-venue competitors where applicable.
4. Present main results first, then ablations, robustness, efficiency, and failure analysis.
5. Explain what each table or figure proves.
6. Separate empirical findings from speculation.
7. Mark missing evidence that should weaken claims.

Rubric

Dimension	Strong	Weak
Claim coverage	Every main claim has evidence	Results do not answer claims
Baselines	Recent and relevant	Convenient or outdated only
Protocol	Datasets/splits/metrics clear	Evaluation setup ambiguous
Ablations	Test key method choices	Cosmetic or missing
Statistics	Variance/significance when needed	Single number overinterpreted
Failure analysis	Honest limits	Only best-case results

Venue Adaptation

• ICLR/NeurIPS/ICML: include ablations, reproducibility details, and robustness/generalization checks.
• CVPR/ECCV: include strong visual baselines, qualitative examples, and failure cases.
• AAAI: make the experimental story understandable across AI subfields.
• TMI: specify modality, cohort, split, external validation, reader/clinical protocol, and privacy constraints where applicable.
• arXiv: include full experimental detail and public artifacts when possible.