These lessons synthesize selected inspected examples. They are editorial guidance, not frequency estimates, official requirements or causal explanations of acceptance. The type guides supply drafting plans, while the atlas reports measured denominators.
Use the contribution's knowledge question to determine section order. A diagnostic paper can progress through competing explanations rather than borrowing a method-paper outline.
Write one question above each results subsection and name the observation that could contradict its proposed answer.
Research-question headings lead from domain specificity to proximity, mixture timing and representation change; boxed findings close each argument. Related work follows the diagnostic sequence rather than interrupting its definitions.
Sections 4–9, pp. 4–9 · Inspected paper
A component-level optimizer table precedes the hypothesis, performance tests, sparsity analysis and explicit limitations; the order makes the control variables legible.
Table 1 and Sections 2-10, pp. 2-9 · Inspected paper
Boundary: Question-led headings are an editorial technique, not an ICML requirement. The examples do not establish how common the technique is.
When execution order or information access changes, draw the relevant state explicitly and connect it to the formal definition.
Use a small trace with named inputs, intermediate state, allowed dependencies and output. Add pseudocode only where the trace leaves iteration or control flow ambiguous.
The paired example and request/completion timeline reuse colors for shared context and independent threads. Stop conditions label each inference step, while the caption distinguishes parallel thread contents from sequential framing tokens.
Figures 2–3, physical p. 3; raster inspected · Inspected paper
Figure 2 aligns token identities with logical positions so physical reordering remains legible; its companion attention mask distinguishes permitted and blocked regions. Figure 5 places the quality–speed frontier beside block-size and model-scale controls.
Figure 2, physical p. 5; Figure 5, physical p. 8; rasters inspected · Inspected paper
Boundary: A diagram explains computation but does not establish correctness or measured speed.
A performance comparison is interpretable only when the budget, evaluation unit and metric are clear. Different fixed budgets can answer different questions.
Put model scale, training exposure, hardware and measurement definition near the result. Separate matched controls from externally reported baselines.
The main protocol fixes token budget rather than raw text volume and openly states that byte exposure differs. Appendix G.2 tests a fixed-text alternative; reporting both makes the practical comparison and its confound inspectable.
Section 3.3, p. 4; Appendix C.3, pp. 22–23; Appendix G.2 and Tables 24–26, p. 40 · Inspected paper
Separates average per-problem speedup, ratio of mean token latencies and measured wall-clock time. Table 3 reports 1.14× wall-clock speedup on 50 MATH problems, while the adjacent protocol discloses one sequential GPU versus four parallel GPUs.
Section 5.2, p. 6; Tables 1–3 and Section 5.5, physical p. 7; raster inspected · Inspected paper
Boundary: Equal tokens, equal bytes, equal FLOPs and equal time are not equivalent controls. Do not claim all of them when only one was fixed.
Separate descriptive association, controlled intervention and formal guarantee. Name regimes where a result changes or fails.
Audit every abstract adjective against the weakest relevant comparison; retain non-winning cells and explicitly scoped limitations.
The timing sweep is explicitly small-scale, and the CKA discussion declines to infer fundamentally different final representations. This keeps interpretation narrower than a suggestive diagnostic plot.
Sections 6–7, p. 7 · Inspected paper
The design-principles discussion explicitly rejects equivalence to bidirectional-attention diffusion models. The causal factorization is a modeling choice whose adequacy is tested empirically.
Section 3.2, physical p. 4 · Inspected paper
Boundary: A careful limitation does not repair an unsupported central claim, and a visually plausible mechanism is not a proof.
Supporting material should answer a reader's concrete verification questions, not simply hold whatever did not fit.
Link each main claim to its proof, exact protocol, data example or robustness result. Use a contents list for long evidence records.
An appendix contents list routes readers to tokenizer definitions, complete questions, perturbation cases and statistical diagnostics. Granular examples explain what a category means rather than supplying only a leaderboard.
Appendix contents, p. 16; Appendices D–H, pp. 23–46 · Inspected paper
Derivations and execution specifications precede stagewise ablations, full worked trajectories and exact rewriting prompts. The separate limitations and future-directions sections keep prospective ideas apart from demonstrated behavior.
Appendices A–H, pp. 12–36 · Inspected paper
Boundary: The inspected published appendix is not evidence that reviewers read every page or that the original submission had identical content.
A position paper needs a contestable thesis and credible alternatives; it may use literature, case studies or new experiments without pretending that all are the same kind of evidence.
State the strongest alternative view, concede what it explains, then identify the remaining disagreement and a concrete implication for practice or research.
Counterarguments receive partial concessions: short-term safeguards and engineering progress are acknowledged, then the thesis is restated as a complementary long-term scientific agenda.
Alternative Views, pp. 8-9 · Inspected paper
Empirical sections correspond to the two opening conditions, followed by counterarguments and four distinct research-agenda pillars.
Sections 3-6, pp.3-9 · Inspected paper
Boundary: ICML 2026 requires Alternative Views for this track. The particular rhetorical sequence and visual package here are recommendations, not additional official rules.
