These patterns synthesize selected inspected examples. They are reusable editorial lessons, not counts of how many papers use each pattern, official requirements or causal explanations of acceptance. Use the contribution-specific plans for a complete drafting package and the atlas for measured denominators.
A specific obstacle, assumption gap or missing capability makes the proposed change understandable before its name is introduced.
Write the conventional formulation and the reason it fails in the target setting. Pair each changed component with that reason; remove a module from the introduction if its purpose cannot yet be explained.
Recasts phase retrieval around the practical difficulty of maintaining an external reference, then promises an analytical self-reference construction rather than an unspecified accuracy gain.
Abstract and Section 1, pp.1–3 · Inspected paper
Uses force as a physical signal for contact rather than only visual proximity, then connects that distinction to more precise supervision.
Abstract and Section 1, pp.1–3 · Inspected paper
Boundary: A compelling motivation is not evidence that the intervention works; it identifies what the experiments or proof must establish.
A diagram should distinguish the data object, changed operation, training signal and deployed inputs. Frozen components and optional outputs affect the claim.
Trace one example through the model. Use distinct labels or line styles for forward computation, supervision and gradients. Put a detail inset around the actual change, not around every inherited component.
The training schematic draws separate expert-supervision paths and the caption explicitly says dense decoding is optional at inference, preventing training-only teachers from appearing required online.
Figure 3, p.5 · Inspected paper
The mask diagram links physical one-second macro-blocks to unequal audio/video token counts and separately displays all four attention directions with labeled query/key axes.
Figure 3, p.7 · Inspected paper
Boundary: A planning diagram is not a reconstruction of every implementation detail. The main figure, equations and actual procedure must agree.
Task scores, resource costs, diagnostic interventions and uncertainty answer different questions. A single rank can hide those differences.
Group comparable settings, retain adverse rows, state metric direction and uncertainty, and name what is held fixed. Separate a selected diagnostic subset from whole-suite performance.
Training ablations test the modules, while inference-time removal or zeroing of geometry tests whether the trained model relies on that input.
Tables 3–5, pp.13–14 · Inspected paper
The vectorization table reports both quality and time, and an ablation without annealing has stronger quality but slower runtime than the full method.
Table 2 and Section 5.3, p.13 · Inspected paper
Boundary: Boldface, a favorable mean or a removal drop does not establish statistical significance, equivalence or an unrestricted causal explanation.
Readers need routes to assumptions, protocol, configurations, full results and failures. Essential meaning belongs in the main argument.
Map main-text claims to named sections of the separate supplement. Keep the defining method and decisive evidence in the main paper; a cited supplement is not automatically inspected in this collection.
Proofs and roughness assumptions are explicitly placed in supplemental material; the main text preserves the propositions and explains their variables. The supplement itself was not inspected.
Section 1, p.2; Proposition 2, p.5; Section 4.1, p.7 · Inspected paper
Hyperparameters, detailed probe architectures and stability metrics are pointed to in supplemental material; these are not counted as separately read sources.
Sections 3.2 and 4.3–4.5, pp.7,11,13 · Inspected paper
Boundary: More supplementary pages do not establish stronger evidence. Public author reports may be longer or differently organized than the accepted revision.
