For a category-level overview before inspecting papers, use writing by contribution type: each category combines corpus measurements, a short synthesis, a section outline and an explicit artifact package. The inline visual gallery shows the forms directly; the 223-paper atlas supplies the quantitative context.
These are editorial lessons drawn from inspected manuscripts, not additional ICLR rules. Each pattern links to concrete paper-level observations and their source locations. The examples span methods, theory, empirical studies, benchmarks and systems; they are selected illustrations, not a statistical ranking of writing styles.
Use this page in three passes:
The complete oral corpus contains the paper-by-paper record. The official requirements govern formatting. The argument worksheet and visual blueprints turn the observations into reusable drafting tools.
Name an obstacle, identify the changed mechanism, and tell the reader which experiment or result will test it. Reuse the same nouns in the contributions, method headings and evaluation questions.
Draft a three-column ledger: limitation → intervention → discriminating test. Delete contribution bullets that merely rename the same intervention.
The introduction repeatedly maps each limitation to one mechanism and one empirical claim: discretization to quality, complex states to tracking, and MIMO to memory-bound decoding. This repeated structure makes a technically dense three-part contribution easier to follow.
Sections 1, 3.1–3.3 and 5 · Inspected paper
Data-mixture and instruction-data ablations are distinct from the cross-task correlation analysis. This separates an intervention from an associative observation.
Table 3 and Figure 3, page 9 · Inspected paper
The introduction narrows the intervention to PPO-like clipping while preserving the surrounding estimator and inference behavior, making the claimed novelty unusually easy to locate.
Sections 1–2, pp. 1–3 · Inspected paper
Boundary: A coherent narrative is not evidence of causality. The intervention still needs an appropriate control, and contradictory results must remain visible.
A worked tensor operation, geometric case or program can introduce the object before the general definition. Reuse its labels through the algorithm and evaluation so each formal step has a recognizable role.
Choose the smallest nontrivial example that exhibits the actual failure. Show its input, one intermediate state, the changed operation and output; explain where the example stops representing the general case.
Code blocks serve as the principal visual device: matched attention implementations use gray panels, colored expressions and corner labels, with identical axis meanings stated once.
Section 5.2, p. 10 · Inspected paper
Numbered callouts and red/blue dashed regions on executable-looking code distinguish recommended decisions from inferred details.
Page 4, Figure 2. · Inspected paper
The Gaussian case establishes the lifting idea before general absolutely continuous measures, with propositions separating cited geometry from new results.
Sections 2–4, pp. 2–7 · Inspected paper
Boundary: A toy example teaches a mechanism; it does not substitute for a proof or establish real-world performance.
Separate the mathematical object, assumptions, guarantee, proof idea and practical algorithm. If a practical approximation does not satisfy the theorem’s conditions, give it its own claim and evaluation.
Place a compact “proved / implemented / observed” ledger beside the main result. State whether the result is finite-sample or asymptotic, necessary or sufficient, and whether optimization attainment is proved.
Two named analytical routes and visibly labelled assumptions let readers distinguish an idealized training plateau from a modified population objective.
Pages 4–7, section 4. · Inspected paper
The progression from counting-friendly languages to arbitrary languages makes the necessity of the encoding phase intelligible; phase I stores an input encoding and phase II simulates a counter machine.
Sections 3–4, pp. 5–9 · Inspected paper
Tables group student scales and distinguish homogeneous from heterogeneous teacher–student pairs. The two compact ablation tables answer component and teacher-choice questions separately. Their very dense page layout is a reminder to inspect tables at actual reading size.
Tables 2–5, main-paper page 9 · Inspected paper
Figure 2 links a global threshold-versus-normalization plot to three fixed-parameter insets; colors distinguish continuous and discontinuous regimes. Figure 3 compares finite dimensions on aligned axes to expose disagreement with the asymptotic transition.
Figures 2–3, pp.7–8 · Inspected paper
Boundary: An accepted theoretical paper can still have an overbroad title, a conditional guarantee or an implementation gap. Do not strip these qualifications when borrowing its presentation.
Different line styles and clearly labeled panels can distinguish training-only labels, teacher signals and privileged observations from deployed inputs. Frozen/trainable status should agree with the algorithm and code.
Trace one test example without looking at the training panel. Every deployed input and operation must be accounted for; no training-only arrow should silently remain necessary.
Figure 1 separates annotation/search, supervised fMRI-to-text training and inference with horizontal bands. Its repeated elephant-and-car example makes the symbolic object descriptions concrete.
Figure 1, page 3 · Inspected paper
Figure 1 consistently assigns red to generator, green to fake model, and blue to frozen teacher; actual images distinguish real/generated inputs. Its caption also states alternating update frequency. Small formulas merit a zoomable version.
Figure 1, page 6. · Inspected paper
The three-stage overview separates alignment, pretraining and frozen inference; its trainable/frozen legend is an effective device for adaptation papers.
Figure 1, p. 3 · Inspected paper
Figure 1 links three views of the same idea: the encoder/predictor graph, trajectories in a maze, and colored latent trajectories whose stars denote predicted barycenters. Matching colors and symbols carry the explanation across panels; the caption explicitly defines the stars, circles and arrows.
Figure 1, page 2 · Inspected paper
Boundary: A dashed arrow has no universal meaning. Include a legend and a caption that explains whether it denotes training only, stop-gradient, optional execution or another relation.
Aligned old/new diagrams, repeated colors and shapes, labeled tensor dimensions, and matched qualitative crops make the changed computation easier to identify than unrelated block drawings.
Keep unchanged modules in the same position and style. Highlight only the intervention; label the feature or state carried by ambiguous arrows. Compare qualitative outputs at matched scale.
The appendix architecture comparison keeps the Mamba-2 and Mamba-3 blocks aligned and uses a shared legend for projections, transformations, optional MIMO operations and nonlinearities. The differences become visible without requiring the reader to compare two unrelated layouts.
Figure 3, page 25 · Inspected paper
The overview uses two clearly separated panels for encoding and generation, with repeated feature-map colors, a visible SMPL condition and tensor dimensions beside intermediate objects. The comparison figure adds magnified crops to normal-map renderings so fine geometric differences can actually be inspected.
Figure 4, page 5; Figure 8, page 8 · Inspected paper
The architecture diagram uses consistent particle colors and separate prior/posterior paths, then zooms into context conditioning. Its caption explicitly distinguishes training-time inverse dynamics from inference-time policy sampling.
Figure 3, p. 7 · Inspected paper
Figure 2 contrasts pairwise attention and core-token processing at the same visual scale. Figure 3 then carries colors and tensor shapes from tokenization through branch averaging and final prediction, clarifying where the module is substituted.
Figures 2–3, pp.4–5 · Inspected paper
Boundary: Color cannot be the only encoding, and an attractive heatmap is not causal evidence. A figure must remain legible in the final PDF at normal size.
Tables should identify data access, model scale, metric direction, aggregation and additional training stages. Separate controlled ablations from literature leaderboard results with unmatched resources.
Write the table caption before bolding any cell. State what is fixed and what differs; mark adapted baselines and values copied from another publication.
The ablation table displays the full coefficient grid, including inferior configurations and infeasible entries, before presenting fine-tuned winners. Curves mark the fine-tuning boundary and teacher/GAN references, connecting mechanism to training behavior.
Tables 1–2 and Figure 2, page 9. · Inspected paper
The matched-token self-distillation ablation isolates whether retaining real data matters. Its explicit no-mixing comparison supports a narrower and stronger conclusion than a general statement that synthetic training avoids collapse.
Appendix F.3 and Table 4, p.34 · Inspected paper
A compact task table exposes important exceptions hidden by domain averages: strict reuse improves some tasks but lowers MMLU, MMLU-Redux and several mathematics/code scores.
Table 2, p. 9. · Inspected paper
Adjacent Tables 3 and 4 reuse task columns while explicitly changing the budget constraint. This makes an efficiency claim auditable without conflating extra processed data with a better architecture.
Tables 3–4, p. 7 · Inspected paper
Boundary: Equal parameter count, equal unique data, equal training tokens and equal inference latency are different controls. A best score is not automatically a fair comparison.
A systems or efficiency claim needs an accounting boundary. Show preprocessing, model execution, auxiliary computation, memory movement and output generation when they affect the user-visible measurement.
Report quality at a fixed resource budget, resource use at a fixed quality target, and measured end-to-end behavior as separate questions when feasible. State hardware, precision, batch, warm-up and included operations.
Accuracy tables group retention ratios and backbone models, while Table 6 separately itemizes vision encoding, compression, LLM forward time and time to first token. Reporting both prefill and end-to-end timing makes clear why a 6.3× prefill gain becomes only about 2.1× at first token.
Tables 1–3 and 6; Section 4.4 · Inspected paper
The appendix distinguishes per-step speed from time to reach a common aesthetic score and reports host RAM separately from GPU memory. Those are more informative comparisons than an isolated memory-saving percentage.
Appendix G, Tables 6–7, page 21 · Inspected paper
The evaluation deliberately separates synthetic expressivity, kernel timing, generation throughput and language quality. It acknowledges failure on the A5 task and Mamba2's lower perplexity, supporting feasibility rather than uniform state-of-the-art superiority.
§5, Tables 1–2, pp.8–10 · Inspected paper
Figure 2 gives a color-coded memory hierarchy and a before/after computation diagram alongside bottleneck shares. It links the implementation change to the resource that is actually scarce.
Figure 2, page 4 · Inspected paper
Boundary: Token reduction and FLOPs are not measured latency; latency and throughput are not interchangeable. A component speedup can disappear behind an unchanged bottleneck.
Benchmark and agent papers must define the behavior being measured, its denominator and the scoring process. A simulator, judge or interaction protocol is part of the scientific instrument, not neutral plumbing.
Include collection/exclusion rules, a scored worked example, judge or simulator checks, leakage controls and failure categories. Distinguish exposure, attempted behavior and successful execution.
Attempt rate and execution-verified success rate distinguish willingness from capability. The manually coded outcome table further separates detection, navigation failure, incomplete execution, and successful attacks, explaining aggregate differences.
Table 1, page 7; Table 3, page 10; Appendix I.6, page 42. · Inspected paper
Figure 3 separates environment resources, priced actions and the user simulator across the top, then contrasts scripted and autonomous control flow below. Reward branches make the evaluation definition visually recoverable.
Figure 3, page 6. · Inspected paper
Table 2 validates the automatic judge against human labels; Table 3 evaluates generators using human labels. Maintaining this separation avoids substituting judge agreement for generator quality, or treating their differently scaled scores as interchangeable.
Tables 2–3, pp.7–8; Tables 5 and 7, pp.15–16 · Inspected paper
Separating instrument validity and reliability from downstream model rankings is a useful benchmark-paper structure.
Sections 4–5, pp. 6–10 · Inspected paper
Boundary: A low score can reflect invalid scoring or navigation failure rather than the intended capability. More tasks do not automatically make the measurement valid.
Use different visual marks for measured points, interpolation, extrapolation and theoretical limits. An explanatory plot must also identify how its variables and normalizations were constructed.
Put the fitting range, uncertainty or sensitivity analysis and any extrapolation boundary in the caption. Keep raw and normalized views when normalization materially changes the story.
Figure 8 distinguishes finite measured models from fitted dashed asymptotes, using arrows to show teacher-to-student compression. Its parameter-count axis makes deployment size visible while leaving training compute a separate question.
Figure 8, p.8 · Inspected paper
Figure 1 places a conceptual decomposition beside training and inference measurements. Its training panel subtracts each dataset-size curve's own minimum; the caption and appendix restore the absolute losses needed to interpret that normalization.
Figure 1, p.2; Figure 15, p.31 · Inspected paper
Three aligned training curves separate radial concentration, AD statistic, and coordinate-wise DP acceptance; the adjacent table states test thresholds and normalization.
Page 9, Figure 4 and Table 1. · Inspected paper
Boundary: A high in-sample fit from a few model scales does not validate a remote asymptote. A qualitative projection is not a complete description of a high-dimensional representation.
Organize proofs, protocols, sensitivity tests, prompts, per-task results and failure cases by the main-text claims they verify. Precise cross-references are more useful than “see appendix for details.”
For each major claim, give a stable appendix destination and the artifact/configuration that reproduces it. Retain version differences and explain which record is authoritative.
The appendix includes negative results, safety degradation, robustness tests, precise prompts and per-asset licenses. This makes the release auditable beyond headline benchmark gains.
Appendices E, I, K, N–P, pp. 30–72 · Inspected paper
A detailed contents page leads to proofs, a runnable-style alternating-update algorithm, implementation settings, and uncurated galleries. The unsuccessful heuristic-SiD extension is retained rather than hidden.
Page 14; Appendix A.3, pages 22–24; Appendices B and D, pages 28–43. · Inspected paper
Appendix F reveals sampling from individual teachers, batch mixing ratios and repeated synthetic data caused by generation limits. These details materially qualify the simplified main-text ensemble-distillation description and are essential for a faithful method summary.
Appendices F.1–F.2, p.33 · Inspected paper
Rejected ambiguous and trivial requests, full prompts and a failed fully automated test-harness attempt expose curation decisions and negative results.
Appendices C–F, pp. 22–27 · Inspected paper
Boundary: Essential validity conditions must remain intelligible in the main paper. ICLR 2026 reviewers were not obliged to read appendices; observed appendix size is not a recommended target.
A negative control, a failed extension or a setting where the baseline wins can identify the method’s actual domain. State the limitation next to the claim it qualifies.
Replace universal adjectives with the tested setting and retain a counterexample or failure analysis. Recheck title, abstract and conclusion after the final tables are fixed.
The appendix supplies dataset links, duplicate counts, attack templates and grids; notably, it reports that ELLie minimal-pair overlap invalidates that dataset for the dilution analysis.
Appendices A, D.2–D.3 and G; pages 25–28, 36–45 and 57. · Inspected paper
Table 1 groups duration and scene-control regimes instead of pooling unlike tasks. Appendix C adds copied-frame and ping-pong baselines that expose how consistency metrics can reward trivial videos.
Table 1 page 7; Table 5 page 17 · Inspected paper
The evaluation table specifies query/target modalities and metrics separately from the training-data table. The common-prompt versus question-specific-prompt comparison directly tests the instruction-awareness claim; the reverse-retrieval appendix exposes a Clotho result below its reference baseline.
Tables 1–2 and 5; Appendix C, Table 8 · Inspected paper
The strongest qualification is quantitative negative transfer: COIN's MoNuSeg PQ falls from0.536 to0.439 after TRACE, and TNBC from0.540 to0.426. The failure figures visibly show mismatched cellular boundaries, turning domain limitations into evidence.
AppendixE.4, Table15 and Figure15, pp.31–32 · Inspected paper
Boundary: Do not infer that a writing choice caused acceptance. This is an oral-only, accepted-paper collection with no matched rejected-paper control; it cannot estimate acceptance effects.
Accepted oral papers are a selected population. Without a suitable comparison group and a design that addresses confounding, these observations cannot show that a particular section order, number of figures, sentence style or appendix length causes acceptance. Scientific importance, soundness, evidence, fit, review judgments and presentation all vary together.
Likewise, a final proceedings manuscript may differ from the anonymous submission that reviewers evaluated. This collection records the inspected version; it does not reconstruct a submission’s revision history without evidence. Learn how a claim is made inspectable, not a supposed formula for acceptance.
