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.
01Keep every guarantee attached to its conditions
A readable theory roadmap should expose where assumptions or update rules change.
Apply it to a draft
Introduce the setting and essential assumptions before the theorem. Use a compact claim-to-result table, local proof intuition and diagnostics that test named claims; label practical algorithm substitutions explicitly.
Lists seven assumptions before presenting the model and Algorithm 1. Theorem 1 retains several terms, and the corollary and following interpretation explain when the interaction term can be absorbed. A practical factorized/Lagrangian optimization is distinguished from the constrained theoretical formulation. This is an assumption-to-algorithm-to-guarantee outline, not an architecture-first outline.
Assumptions 1–7, pp.3–4; Equations 9–16, Algorithm 1, Theorem 1 and Corollary 1, p.5; interpretation, p.6 · Inspected paper
Uses explicit questions and phase names to make a technical analysis navigable, and ends with a direct limitations paragraph. Crucially, the Phase II lemmas analyze gradient descent for tractability after entering a small-weight regime, while the SAM section offers empirical analogies. Those update-rule and epistemic changes must remain visible in a summary; the two-phase table alone could otherwise suggest a stronger continuous noisy-SGD guarantee than the body establishes.
Phase II and Lemmas 5–6, pp.5–6; Extension, pp.6–7; Limitations, p.7 · Inspected paper
Boundary: A two-phase narrative is not automatically an end-to-end theorem, and an empirical extension does not inherit the original guarantee.
02Organize analysis around controlled questions
An empirical investigation needs competing explanations and protocol-aware comparisons, not necessarily a new architecture.
Apply it to a draft
Name the question, what varies, what stays fixed and the sampling unit. Use matched diagnostic panels or resource sweeps, then explain how the evidence motivates any follow-up intervention. Put executable filtering and prompt details in a versioned appendix.
Uses question-driven subsections and local discussion blocks to connect an experimental manipulation, a result and an interpretation. Within planning, the order is turns, reasoning length, difficulty, then diversity, before a practical follow-up. A reader can find the comparison settings before the conclusions. However, prompt-only analyses and LoRA analyses have different protocols and sample budgets, so adjacent tables are not a single controlled ranking of three intrinsic capabilities.
Sections 3–4, proceedings pp.2–6; extended Appendix A, pp.10–11 · Inspected paper
The extended version provides operational material rather than only additional scores: Appendix A lists data sources and analysis settings; B explains model-assisted taxonomy consolidation; C gives executable-style filtering pseudocode, medium-turn/difficulty criteria, model-specific final-training settings and prompt artifacts. This supports a claim-indexed appendix plan. Its full file repeats main text and references, so the supporting-material span must start at Appendix A rather than treating the whole 13-page author PDF as an appendix.
Boundary: Sampled failure labels are not a causal decomposition; adjacent tables with different budgets are not one controlled ranking. Preserve unresolved inconsistencies between prose and figures.
03Define what success means in the application
A domain result needs a domain-specific information and evaluation contract.
Apply it to a draft
Show the acquisition or conditioning constraint, separate surrogate metrics from actual task success, and retain raw trial counts or absolute recovery where they matter.
The main comparison separates MIST and BCI and gives directions for FID, KID, SSIM and DAB histogram divergence. Table 2 is a two-by-two component ablation: the full method improves several measures, but DBD-only has higher SSIM (0.251 versus 0.248). The text itself explains why SSIM is limited under weak pairing. The inspected tables do not provide repeated-run uncertainty; selected tissue images and distributional image metrics do not establish diagnostic performance.
Evaluation Metrics, p.5; Tables 1–2 and Ablation Study, p.6; Tables 3–4, p.7 · Inspected paper
Table 1 groups two datasets and separates exact top-k recovery, Tanimoto similarity and graph-edit distance. The setup reports official splits, a pretraining/test molecular-distance exclusion and eight random seeds. Table 2 tests the input-information contract: its label 'w/o Formula' actually means formulas predicted by MIST-CF instead of ground-truth formulas. MSAnchor's top-10 similarity on MassSpecGym is 0.41 in Table 1 but 0.42 in Table 2; the discrepancy is unresolved and must not be silently harmonized.
Table 1 and Datasets/Metrics, p.5; Table 2 and Ablation on Molecular Formula, p.7 · Inspected paper
Table 1 reports task counts and simulation success by category; Table 2 retains raw real-robot successes/trials and a missing entry. The full model loses to its no-foresight ablation in Scenario II (11/40 versus 13/40). Generated-goal Table 3 concerns separately trained Scenario III policies, not every robot result. The headline 41% improvement corresponds to a 41-percentage-point simulation gap (86% versus 45%), and should not be rewritten as a relative gain or universal deployment result.
Evaluation protocol, pp.5–6; Tables 1–3 and VLM-FoAM Joint Inference continuation, p.7; Abstract, p.1 · Inspected paper
Boundary: Attractive tissue images are not diagnostic validation, molecular similarity is not laboratory confirmation, and simulation success is not an unrestricted robot-deployment result.
04Turn a general motivation into a testable gap
Name the setting where the familiar approach fails before introducing its replacement.
Apply it to a draft
Use a controlled example to connect the failure to the changed computation, then repeat those nouns in the contribution list and experimental questions.
The opening narrows a broad efficiency problem to a missing setting: image-token pruning inside multimodal in-context sequences. Figures 1–3 make the failure case visible before the new method, linking the contribution to cross-image/text interactions rather than merely claiming fewer tokens.
The opening connects a familiar technique, weight averaging, to a specific failure under unequal modality learning speeds. Its contributions follow diagnosis → intervention → evaluation, giving the reader a reason for each component rather than presenting a module list without a problem.
Abstract and Introduction, physical pp.1–2; Figure 1 · Inspected paper
Boundary: A compelling example does not establish prevalence or prove the proposed mechanism.
05Show computation and supervision separately
An architecture figure should explain ownership, state and information flow.
Apply it to a draft
Give the overview one job, zoom into the nonstandard operation, and label teacher updates, training signals, pruning locations or inference-only paths.
Figure 2 uses a top-level student/EMA-teacher view and a lower zoom into dominance scoring and dropout. Encoders, features, heads and losses have a legend, while arrow routing exposes the supervisory relationship. Figure 3 answers a different question through perturbation-response plots; it is not another architecture diagram.
Figure 4 follows token movement vertically through projector and decoder layers, using token colors and pruning marks to expose the stage boundary. It gives architecture placement, while Figure 5 examines context configuration and Figure 6 examines hyperparameters. The very short architecture caption leaves important interpretation in adjacent prose; a self-contained caption would help reuse.
Figure 2 aligns layer-index panels across two models and two benchmarks, allowing the reader to compare entropy and two retrieval constructions. Figure 3 uses a branching token diagram: contextualized features feed retrieval, unselected visual-page tokens are removed, and retained visual tokens plus query tokens continue into later LLM layers. The visual distinction between retained and discarded token blocks is the explanatory job of the architecture figure.
The three named method subsections follow the same cycle shown in Figure 2: estimate flow using frozen foundation priors plus task adaptation, update depth/poses with two residual types, and use the resulting reliability information in the next refinement. Algorithm 1 supplies the conditional update logic that the architecture picture alone cannot specify.
Sections 3.1-3.3 and Figure 2, pp.2-4; Algorithm 1, p.4 · Inspected paper
Boundary: A colorful block diagram without operation labels is not a method explanation; avoid copying an accepted paper’s unexplained arrows.
06Keep the comparison contract beside the numbers
A result is interpretable only with task, input access, split and measurement unit.
Apply it to a draft
Separate training-domain and test-domain directions, model budgets and modality access. Report variation with its meaning and preserve baseline wins.
Tables 1–2 name the source-to-target direction and modality access and report per-split mean and standard deviation over three seeds. Table 4 isolates EMA and selected added components; Table 5 compares fixed and adaptive dropout. Retain non-wins: the HAC video+audio average in Table 1 is 70.66 for MBCD versus 71.25 for CMRF.
Table 3 keeps latency, FLOPs, KV-cache memory and relative task performance separate. In its LLaVA-Next-7B block, DivPrune reduces FLOPs more than CATP but has higher latency than vanilla, illustrating why a proxy cost cannot substitute for a measured endpoint. The caption says four models, but this published table visibly contains three model blocks; preserve that source mismatch.
Boundary: A source may contain mismatched captions or missing conditions; retain the discrepancy instead of polishing it into a stronger claim.
08Keep the core argument in the main paper
Support should make the claim verifiable rather than conceal its definition.
Apply it to a draft
Put essential assumptions and operations in the body. Point to exact proof, hyperparameter, selection and data details in any optional support, and verify the target exists.
The publisher version supplies the central losses and interpretation in the main body and a code link on its first page, then ends with references rather than an attached appendix. A future drafting plan should place exact hyperparameters, model-selection rules and full implementation details in accessible support; that recommendation is not a claim that such support was inspected here.
Code link p.1; Eqs.5–9 p.4; references pp.8–9 · Inspected paper
The accepted PDF carries the core selection equations and settings in its main body, and links an extended version separately. These observations and counts use only the nine-page publisher file. Do not infer that the extended version lacks proofs, extra experiments or pseudocode merely because they are absent here.
Extended-version link p.1; Eqs.1–10 pp.3–5; end pp.8–9 · Inspected paper
Boundary: An extended-version link is not an inspected appendix; a poster is not supplementary proof.
09Learn useful patterns without inheriting overclaiming
Accepted examples are evidence to inspect, not perfect prose to imitate.
Apply it to a draft
Distinguish observation from mechanism, empirical generalization from a theorem, and the tested subset from a universal result.
The strongest reusable convention is the repeated mapping from imbalance to a named corrective operation. Do not imitate the paper’s broad superiority language without qualifications: average gains do not establish every-split wins, and a t-SNE picture or local loss-flatness diagnostic does not independently prove the asserted mechanism.
Introduction p.2; Table 1 p.5; Further Analysis and Conclusion pp.6–7 · Inspected paper
Terms for demonstrations and the query make the input contract concrete, while bold local signposts help the reader distinguish design decisions. Calibrate the conclusion to the tested setting: the relative aggregate is not absolute accuracy or a win on every benchmark, and “first” language still requires a defensible novelty search.
