Argument and evidence worksheet
This is an editorial worksheet, not a conference form. Use it before drafting prose. Its purpose is to expose missing evidence while there is still time to improve the study.
Download the editable Markdown worksheet. It is an original planning aid; obtain formatting files from the conference’s official template.
1. Write the claim before the abstract
Section titled “1. Write the claim before the abstract”Complete one sentence:
In [defined setting], changing [specific mechanism or assumption] produces [measured or proved effect], compared with [appropriate alternative], under [important constraints].
Replace adjectives with conditions. “Efficient,” “general,” “robust” and “interpretable” need operational definitions. A paper can have several contributions, but the reader should know which is central and which supports it.
| Vague claim | More inspectable claim shape | Evidence it would require |
|---|---|---|
| Our model is efficient | Inference latency falls at a fixed quality target on a named device and workload | Timed protocol, batch size, warm-up, precision, memory, quality and variance |
| Low magnification matches high magnification | Under a specified field-of-view and patient split, low-magnification-only inference reaches a defined comparison target | Acquisition details, paired sampling, training privileges, test-time inputs, uncertainty and external tests |
| The model reasons better | A defined reasoning failure decreases under controlled prompts and inference budgets | Task validity, contamination checks, error categories and matched budget |
| The representation is more interpretable | A specified intervention or explanation criterion improves | Target definition, intervention/control, reliability and counterexamples |
| The method is theoretically justified | A stated result holds under explicit assumptions | Definitions, theorem, proof, dependencies and scope of applicability |
These are claim templates, not results from a particular paper. Do not fill a number or success statement until the evidence exists.
2. Build a claim-to-evidence ledger
Section titled “2. Build a claim-to-evidence ledger”Keep this table beside the draft. Every major abstract or conclusion claim should have a row.
| Field | What to record |
|---|---|
| Claim ID | A stable short name, such as C1: matched-budget quality |
| Exact proposition | What changes, in which setting, relative to which alternative |
| Supporting item | Experiment, theorem, figure, table or release artifact |
| Controls | Data, split, architecture, initialization, supervision, tuning and compute held fixed |
| Alternative explanation | The strongest competing interpretation of the result |
| Discriminating test | An ablation, negative control, counterexample or sensitivity test |
| Uncertainty | Repeated runs, confidence interval, statistical assumptions, measurement noise or proof conditions |
| Boundary | The population, regime or capability not established |
| Manuscript location | Where the claim is stated and where its evidence is explained |
| Status | Supported, partly supported, contradicted or not yet tested |
Do not turn an unsupported row green by moving the caveat to the appendix. Either acquire the missing evidence or narrow the claim.
3. Identify the contribution type
Section titled “3. Identify the contribution type”Method or architecture
Section titled “Method or architecture”Explain the old bottleneck, the changed computation, and why that change addresses the bottleneck. Trace one example through inputs, representations, objectives and outputs. Separate training-only supervision from deployed computation. An architecture picture should expose the new mechanism, not simply give every module a colored box.
For distillation, identify the teacher, student, transfer signal, alignment, loss weights and inference contract. State whether the contribution is a new target, alignment, optimization schedule, privileged input or deployment trade-off. Avoid calling all teacher–student learning “residual KD.”
Empirical finding
Section titled “Empirical finding”State the question, the population of models/tasks examined and the intervention or comparison. Explain how the protocol distinguishes a real phenomenon from selection effects. Report important counterexamples. A broad title should not rest on one model and one benchmark without an explicit scope boundary.
Theory
Section titled “Theory”Give the reader the objects, assumptions and result before the proof machinery. Explain the theorem in ordinary language, then state it precisely. Separate sufficient from necessary conditions, worst-case from typical behavior, finite-sample from asymptotic claims, and proved conclusions from experiments that merely illustrate them.
Benchmark or dataset
Section titled “Benchmark or dataset”Define what the resource measures and why existing measurements are inadequate. Explain sampling, annotations, provenance, splits, exclusions and access terms. Test whether the benchmark measures the intended capability rather than artifacts or leakage. A difficult benchmark is not automatically a valid one.
System
Section titled “System”Specify the workload and operating constraints, then explain the design choices that meet them. Include the data/control flow, interfaces, failure handling and reproducible performance protocol. Separate algorithmic savings, implementation optimization and hardware differences.
4. Draft the introduction as a chain of obligations
Section titled “4. Draft the introduction as a chain of obligations”One useful—but not mandatory—progression is:
- Setting and stakes: what researchers or users need to do.
- Specific obstacle: what prevents the existing approach from doing it reliably.
- Gap in the closest work: which assumption, mechanism or evidence is missing.
- Insight: the observation that motivates your design or analysis.
- Contribution: what you actually provide.
- Evidence preview: which tests or results substantiate it, with scope.
Each transition should answer a reader’s next question. “Deep learning has achieved remarkable success” rarely identifies the particular problem. Start close enough to the actual scientific tension that the reader knows why this paper exists.
If you use a contribution list, make each item testable and non-overlapping. “We propose a method,” “we design an architecture,” and “we introduce a framework” can be three names for one contribution. A resource release and a verified scientific finding may instead be distinct contributions.
5. Design comparisons before presenting scores
Section titled “5. Design comparisons before presenting scores”For each table, write down the comparison contract:
- Unit: image, video, sequence, patient, slide, case, task or model.
- Data: exact datasets, versions, exclusions, split construction and overlap risks.
- Access: labels, teachers, external data, prompts, demonstrations and privileged information.
- Capacity: parameters, active parameters, tokens, context and backbone.
- Budget: training, tuning and inference compute—not only one favorable component.
- Selection: which validation data choose the checkpoint or configuration.
- Measurement: metric definition, direction, aggregation and uncertainty.
An ablation should test a mechanistic hypothesis. Removing a component and retraining with a very different optimization budget can confound the interpretation. A fair comparison does not always mean identical hyperparameters; explain the tuning protocol and whether each baseline had a reasonable opportunity to perform well.
Example: a cross-magnification distillation paper
Section titled “Example: a cross-magnification distillation paper”An original planning ledger might separate four questions:
| Question | Comparison | Essential qualifier |
|---|---|---|
| Is the low-magnification student competitive? | Student against its high-magnification teacher and a same-input baseline | Test-time access, optical scale/MPP, field of view, patient split |
| Does the transfer mechanism help? | Same student with and without the proposed transfer signal | Matched data, backbone, training and tuning budget |
| Is a residual component necessary? | Residual formulation against a direct target with a controlled parameter budget | What “residual” means mathematically and where it is computed |
| Does the efficiency claim survive deployment? | End-to-end latency and memory at a stated operating point | Tissue detection, extraction, encoding, aggregation and remaining high-resolution stages |
Do not imply that this worksheet is a verified result, a required ICLR protocol or a clinical validation standard.
6. Write the abstract last enough to be truthful
Section titled “6. Write the abstract last enough to be truthful”A useful drafting sequence is problem → gap → idea → evidence → bounded implication. It is not a required sentence count. Make the subject of each sentence clear. Name the novel operation, not merely the method acronym.
Check that:
- the abstract’s strongest claim is supported in the main text;
- important units and comparison baselines are identifiable;
- percentages distinguish relative change from percentage-point change;
- a theoretical result retains its key assumptions;
- a resource contribution identifies what is released;
- the conclusion does not expand beyond the evaluated setting.
An abstract should make the reader want to inspect the evidence, not require trust in a string of superlatives.
7. Calibrate language without making it vague
Section titled “7. Calibrate language without making it vague”| If the evidence is… | Prefer a statement that… | Avoid… |
|---|---|---|
| An observed association | Describes the measured relationship and study conditions | Causal language without an identifying design |
| A controlled ablation | States which competing explanation the intervention weakens | Claiming the only possible mechanism has been proved |
| A theorem | Preserves assumptions, quantifiers and type of guarantee | Generalizing the theorem to an unmodeled architecture |
| Several successful benchmarks | Names the tested tasks and meaningful exceptions | “Universally superior” or “solves reasoning” |
| A plausible transfer to pathology | Labels it as a hypothesis requiring WSI-specific validation | Implying natural-image results already demonstrate clinical utility |
Hedging is not a substitute for scope. “May significantly improve generalization” is less precise than stating the observed effect on a defined external cohort.
8. Run three revision passes
Section titled “8. Run three revision passes”Argument pass: can a reader reconstruct the claim–mechanism–evidence chain from the introduction, main figure and results? Does every major claim have a support location?
Evidence pass: verify every number, citation, split description, baseline label, unit and inference-cost statement against the actual experiment or proof. Check disagreement between prose and tables instead of silently choosing the more favorable number.
Reader pass: remove undefined acronyms, inconsistent notation, ambiguous pronouns, unnecessary nesting and repeated claims. Keep the nearest relevant explanation beside the equation or figure. Ask a colleague outside the immediate project to explain the mechanism back to you.
Finish with the structure and visual-design preflight and the rules for the exact conference edition. A completed checklist documents careful preparation; it does not certify novelty, correctness or acceptance.