Choose the route closest to your main contribution, then adapt it. Architecture, objective, theory, analysis, resource, system, application and position arguments carry different evidence obligations. Only types present in this inspected corpus appear below. These editorial routes do not replace the conference's official track definitions.
For a new design, prepare a readable computation or representation overview, a comparison with matched conditions, and controls that test the changed component. Add pseudocode when iteration, state, sampling or gradient routing would otherwise be ambiguous. Add a theorem only when making a formal claim. A position paper instead needs a defensible thesis, substantive alternatives and implications; it is not required to pretend to be a leaderboard method.
Observed medians describe source versions. Suggested display packages are conditional editorial starting points, not official requirements or evidence of a necessary number of figures. Small groups provide examples, not robust estimates of a conference-wide standard.
The central deliverable changes the model's representation, component arrangement, information flow or reusable inference pipeline. Data, objectives and efficiency can still be substantial supporting contributions.
Corpus context · 22 papers: median 6 main figures and 4 main tables; body span ≈ 9.11 pages; attached appendix span ≈ 12.91 pages (21 resolved boundaries). Separate source files: median Not resolved whole-file pages among 0 papers; these can repeat main text and are not appendix-only lengths. Extraction-assisted observations—not drafting targets.
Make the representation change legible before the leaderboard, then test its claimed mechanism separately from additional capacity. Name the obstacle, show the intermediate representation or symmetry operation, and include a matched-capacity or matched-operator control. Exact discrete symmetry, empirical robustness, effective rank and downstream accuracy support different claims. These are lessons from the named examples, not frequency estimates for every member.
Editorial starting package: choose artifacts that test your actual claims. Quantities below are planning suggestions, not counts inferred from all category members or conference rules.
Start with one overview; add insets only when they resolve different questions
Label inputs, outputs, inherited versus new components, frozen/trainable blocks and training-only paths.
When: Core when the novelty is computational organization; use a simple representation cartoon before the full diagram if the data object is unfamiliar.
One compact example or aligned qualitative grid
Make the claimed representational problem visible and reuse the same example through the mechanism.
When: Especially useful for 3D geometry, multimodal tokens, articulated objects or structured outputs; do not add a gallery that does not test the claim.
One focused table per incompatible task or information regime
Keep metric direction, data, backbone, reference inputs and refinement policy next to the results.
When: Separate single-model inference from per-instance optimization and privileged-input baselines.
A compact switch table or matched diagnostic panels
Test the named reason for each component rather than removing modules without preserving compute or capacity.
When: Use matched controls when the mechanism claim depends on architecture rather than more parameters or more supervision.
Only when the procedure or guarantee requires it
Expose iteration, state, branching or assumptions not recoverable from the diagram.
When: A feed-forward module does not require pseudocode or a theorem simply to look like an accepted paper.
Layer and tensor specifications, training/data recipes, additional task-level results, ablation settings, input/refinement contracts, runtime definitions and failure cases. Keep the defining representation and actual changed computation intelligible in the main text.
Use a shared algebraic baseline map and targeted controls to separate added capacity from the proposed update geometry.
Map each architectural change to a stated invariance failure and evaluate the claimed symmetry separately from task accuracy.
Boundary: This route contains 22 primary editorial assignments in the 168-paper inspected subset (168 inventory records). It is not an official track, a universal writing standard or evidence that these choices caused acceptance.
See inline examples of these artifact styles · Compare every member in the measurement ledger
The primary change is a recursively compressed, tokenizer-free genomic representation with learned chunking; training and scaling tests support that architecture.
A next-scale autoregressive representation and shared flow decoder reorganize protein backbone generation; exposure-bias and sampling changes support that structure.
A pretrained sequence-to-sequence operator for scores and densities is the primary deliverable; formal characterization and estimation tasks support the architecture.
The central change augments the state-space recurrence with momentum and conditioned writes while preserving parallel scans.
The contribution redesigns tokenization, coordinate transitions and decoding around a reference-basis hierarchy while retaining standard attention.
A polynomial policy parameterization is the transferable method; Mountain Car analysis supplies an interpretable reference for evaluating that representation.
The central contribution is a heterogeneous routed expert architecture for unified multi-dataset, multi-modal detection.
The contribution is an event-cloud architecture with polarity-aware grouping and frequency-domain feature extraction.
The central method introduces a behavior encoder, phase-conditioned decoder and retrieval-based representation architecture.
The central contribution is a dual-branch discrete representation architecture integrated into a three-stage policy pipeline.
The central contribution is a hierarchical latent-hypergraph model and structured inference architecture for shared disease-risk effects.
The main contribution changes neural-operator architecture through joint state/mesh evolution and error-driven latent updates.
The central contribution is a dual-stream latent architecture coupled through a differentiable tangent mapping.
The main contribution adds rational recurrent features and repeated layer-wise injection to a Transformer architecture.
The contribution combines variate ordering, patch-level attention and a flexible Fourier-basis decoder in a multivariate forecasting architecture.
The method introduces a discrete semantic bottleneck and confidence-gated conditioning architecture for text-informed forecasting.
The method changes amortized-sampler architecture through a learnable latent dynamical state and recurrent policy parameterization.
A unified autoencoder and quantized latent workflow change where weather reconstruction and downstream prediction occur; data release and compression measurements support that mechanism.
The intervention is a client-specific geometric representation and parameter-decoupling/aggregation design.
A plug-in side module injects geometric features into a frozen VLM; geometric losses and task datasets train and evaluate this architectural intervention.
The primary design separates a learned discrete theory programmer from a reusable primitive executor; grounding and minimum-description-length losses train this compositional architecture.
The principal contribution is a modular learned projection layer and guided architectural variant for privacy-aware anomaly representations.
The argument primarily changes a loss, supervision signal, optimization/update rule, distillation recipe or adaptation procedure. Some assigned methods update at inference time; the label refers to the objective-driven argument, not only offline training.
Corpus context · 35 papers: median 6 main figures and 3 main tables; body span ≈ 8.83 pages; attached appendix span ≈ 11.91 pages (35 resolved boundaries). Separate source files: median Not resolved whole-file pages among 0 papers; these can repeat main text and are not appendix-only lengths. Extraction-assisted observations—not drafting targets.
Keep the ideal objective, implemented estimator and behavioral failure mode visibly separate. Trace the diagnosed problem into the objective and estimator, disclose approximation or skipping rules, and test the intervention under a matched evaluation contract. An objective-derived limit need not describe finite-compute implementation; eliciting a prior is not injecting missing knowledge. These are lessons from the named examples, not frequency estimates for every member.
Editorial starting package: choose artifacts that test your actual claims. Quantities below are planning suggestions, not counts inferred from all category members or conference rules.
One diagram when multiple paths or update stages matter
Mark teachers, students, pseudo-labels, frozen weights, detached gradients and inference-only operations.
When: Use gradient arrows or phase panels when an ordinary module diagram would hide the actual intervention.
An explicit objective; an algorithm box if iterative or stateful
Define variables, positive/negative samples, weights, updates and stopping rather than listing an unexplained weighted sum.
When: A short closed-form loss may need no pseudocode; a multistage update usually does.
One task-specific table or paired baseline/new-rule rows
Hold backbone, supervision and budget fixed where possible so a training-rule claim is not merely a larger-model claim.
When: Show unmatched leaderboard context separately from causal design comparisons.
One appropriate panel per central explanatory claim
Measure the suspected mismatch: gradient behavior, confusing classes, reliability, alignment or temporal persistence.
When: Use the quantity the paper actually claims to improve, not an arbitrary feature visualization.
A compact control table and consequential parameter curves
Distinguish new supervision, loss terms, schedules and extra compute; retain collapse or negative runs.
When: Do not invent large sweeps for inconsequential parameters or hide tuning costs.
Optional
State assumptions and whether the implemented approximation satisfies them.
When: A proof is needed for a formal claim, not as decoration for an empirical improvement.
Exact objectives and update order, full derivations where used, sampling and teacher/data provenance, schedules, tuning budget, seed-level results, alternative objectives, cost and sensitivity. Keep the motivation for each main loss term in the main paper.
Trace a training objective all the way to the exact implemented estimator, explicitly recording any order, bias or protocol change introduced for practicality.
Make the hypothesis-to-intervention chain explicit and include a task where the intended inductive bias predictably fails.
Boundary: This route contains 35 primary editorial assignments in the 168-paper inspected subset (168 inventory records). It is not an official track, a universal writing standard or evidence that these choices caused acceptance.
See inline examples of these artifact styles · Compare every member in the measurement ledger
The central change learns an unmasking policy using a correctness-speed reward, leaving the underlying diffusion language model fixed.
Feature-coverage-guided synthetic-data construction is an objective-driven post-training recipe; its metric and analysis support that procedure.
Online data selection changes the optimizer-aware training update, with efficiency machinery supporting practical use of that selection objective.
The central argument is a mid-training recipe organized around agent-native supervision; released corpora support that training intervention.
Decoupled objectives and coordinated update timescales are the organizing intervention, with memory blocks implementing the optimization scheme.
The central contribution is a sample-wise denoising-style metric objective; formal guarantees and geometric applications establish its target and usefulness.
Motion-weighted gradient attribution drives targeted fine-tuning data selection; efficiency approximations support that training-data intervention.
A constrained optimizer update and weight geometry are primary; distributed infrastructure makes that optimization intervention practical.
A bounded token loss is the primary intervention; diagnosis and derivations motivate its selective suppression of low-probability supervision.
The primary contribution is a compute-indexed correctness objective and implementable advantage normalization; theory explains which objective the estimator actually optimizes.
The contribution changes policy-parameter updates through a Lyapunov-constrained projection; human-robot collaboration supplies its evaluation setting.
Adaptive partner generation and learnability-based curriculum selection change the training distribution without requiring a pretrained partner population.
The contribution constructs a timestep-aware error-compensation procedure motivated by quantization-error propagation.
The contribution changes reward modeling and its training/inference construction to mitigate preference-learning failure modes.
The primary methodological claim is reinforcement learning with search-grounded rubrics that evolve alongside the policy.
The methodological change uses intermediate translation rewards to learn a latency-quality policy without fine-grained alignment.
The contribution builds a semi-supervised reward-learning procedure using optimal-transport pseudo-preferences.
The method changes reward verification and token-level credit routing in multimodal reinforcement learning.
The central solver combines measurement fitting with entropy and structural regularization in a randomized compression objective.
The contribution derives and tests optimization updates based on posterior correction.
The contribution changes quantization-aware supervision and initialization to preserve reasoning behavior.
The central contribution modifies the multi-objective preference update with gradient-conflict handling and clipping.
The method introduces conditional-dependence estimators and a differentiable penalty for informed bias mitigation.
The contribution couples a margin-prioritized curriculum and contrastive robust-distillation objective, with an efficient inner attacker supporting the training recipe.
A student-conditioned local scoring rule changes response selection for supervised distillation; diagnosis and ablations justify that training-data intervention.
A distributional reward objective and risk-sensitive policy surrogate are the central intervention; theory and three experimental settings test their behavior.
Three critic-side regularizers implement a geometry-based smoothness hypothesis; the paper links the surrogate losses to measured field behavior and control trade-offs.
A structured preference reward and gated training signal are the principal interventions, supported by an explicit trace-quality framework.
The central intervention coordinates local contrastive discrimination with a global hierarchical energy objective, supported by spectral architecture and conditional theory.
The primary intervention is staged supervision and RL that teaches an existing VLM an orthographic intermediate representation; the new benchmark diagnoses and evaluates that training design.
The central contribution trains a shared solver/verifier through internal process rewards and iterative policy updates; external tools and role prompts implement the learning loop.
The main intervention changes the pretraining curriculum and transferred initialization using algorithmically generated nonsemantic data; diagnostic experiments investigate why and where this intervention transfers.
The central intervention is supervised, gradient-free optimization of prompts and collaboration controllers across functional and structural dimensions, rather than a new base architecture.
The main intervention is a joint diffusion/signature training objective that conditions output behavior on authorization.
The contribution is a closed-form constrained parameter-editing objective for targeted concept removal and retention.
A formal characterization, algebraic result, guarantee or theoretically derived explanation organizes the main argument, with algorithms and experiments testing consequences or practical approximations.
Corpus context · 38 papers: median 3.5 main figures and 0.5 main tables; body span ≈ 9.05 pages; attached appendix span ≈ 18.88 pages (37 resolved boundaries). Separate source files: median Not resolved whole-file pages among 0 papers; these can repeat main text and are not appendix-only lengths. Extraction-assisted observations—not drafting targets.
State the comparison contract before presenting an upper/lower-bound contrast. Define input domain, computational model, access, rounding or delay assumptions alongside each theorem; use the appendix as an explicit proof map. Statistical optimality need not imply efficient computation, and a finite-precision expressivity theorem is not a training guarantee. These are lessons from the named examples, not frequency estimates for every member.
Editorial starting package: choose artifacts that test your actual claims. Quantities below are planning suggestions, not counts inferred from all category members or conference rules.
One when it makes unfamiliar definitions operational
Connect algebraic symbols to camera geometry, representations or update behavior.
When: Prefer a reusable symbolic example over an unrelated network schematic.
One short map or comparison table when several regimes are involved
Keep quantifiers, non-degeneracy conditions, approximation regimes and guarantee types visible.
When: Especially important when exact theory motivates an approximate practical solver.
As many as the actual logical claims need
State what is proved and identify the central argument before full technical details.
When: No target theorem count; empirical observations must not be promoted to universal guarantees.
Only when there is an empirical consequence to test
Test the predicted difficult regime, scaling, geometry or optimizer behavior under stated assumptions.
When: A finite experiment illustrates or challenges a consequence; it does not prove an asymptotic statement.
Include where needed to prevent a tempting overgeneralization
Explain the role of assumptions and distinguish sufficient from necessary conditions.
When: A theoretical paper need not manufacture a broad benchmark leaderboard if it is irrelevant to the result.
Section-matched proofs, notation, technical lemmas, numerical construction details, exact solver settings and differences between the theorem model and the experiment. Main text must retain essential assumptions and the meaning of the main result.
Put the computational semantics and domains in the comparison table's contract, and distinguish a tight result from a remaining upper/lower gap.
Explain what each bound term costs, and keep statistical optimality, computational hardness and practical runtime as distinct claims.
Boundary: This route contains 38 primary editorial assignments in the 168-paper inspected subset (168 inventory records). It is not an official track, a universal writing standard or evidence that these choices caused acceptance.
See inline examples of these artifact styles · Compare every member in the measurement ledger
The core argument translates memory criteria into algebraic constraints and a closed-form decomposition, with empirical manipulations testing that formalization.
Formal optimality under a defined logical search space is the central claim; empirical search cost and explanation differences demonstrate its practical consequence.
Equilibrium invariance and last-iterate convergence theorems organize the contribution; game experiments illustrate the formal intervention and its extension.
Moment-SOS formulations and convergence/certification guarantees are primary; clustering is explicitly an illustrative application.
A non-asymptotic convergence bound under relaxed assumptions is the main deliverable; learning-rate experiments illustrate consequences without replacing the theorem.
Finite-round revenue guarantees derived through approximate equilibrium and dual fitting are the core contribution.
An algebraic solver family and its reversibility/convergence relationships organize the paper; applications test those mathematical properties in finite precision.
Matching statistical regret bounds, a lower-bound construction and a general transition abstraction are the main evidence, without an empirical benchmark claim.
The central contribution characterizes minimax decisions and population-optimal prediction sets; the finite-sample algorithm and experiments instantiate that characterization.
Definitions, layer characterizations and distinct universality theorems are the contribution; architecture examples illustrate the mathematical scope.
Upper and lower representation bounds under explicitly defined floating-point arithmetic are the entire evidentiary contribution, not a training or hardware benchmark.
Packing-number upper bounds under finite precision provide the core claim; empirical cramming and geometric refinements examine their relation to observed behavior.
The central deliverable is a convergence and accuracy-complexity guarantee for a score-based sampler.
The core deliverable is a mechanism with formal fairness and truthfulness properties under explicit assumptions.
The paper’s central results are a formal incentive model, hardness result and characterization of incentive-compatible additive pricing.
The core result is an explicit mathematical decomposition and its associated linear-system representation.
The main contribution analyzes convergence and timescale structure of a joint gradient flow.
The main result is regret and constraint-violation analysis for causally constrained contextual bandits.
The central contribution establishes recursive-decomposition and reconstruction guarantees for latent-variable causal discovery.
The main contribution is a Trie-based representation and local identifiability theory for count-data causal models.
The primary contribution develops a choice-theoretic and proper-loss framework that characterizes preference-optimization designs.
The main contribution is a first-order rejection-sampling construction with complexity guarantees for diffusion and log-concave sampling.
The primary contribution proves delayed generalization and hyperparameter-dependent timing bounds in ridge regression.
The core contribution is an auxiliary-variable MCMC framework with correctness, ordering and transition-comparison theory and derived samplers.
Compute-bounded policy formalization and existence results organize the contribution; a minimal recurrent architecture supplies empirical tests rather than an all-purpose scaling guarantee.
The central evidence is a curvature-dependent oracle-complexity argument and constructive estimator; the manuscript is not an empirical benchmark paper.
The primary contribution derives necessary representational geometry and dimension constraints under explicit compositional-generalization assumptions.
A norm-dependent mathematical definition and conditional stability analysis unify the empirical optimization diagnostics.
The core claim is an intrinsic randomized mechanism with formal privacy, invariance and noise-energy characterization, supported by downstream utility diagnostics.
The main contribution is an ambiguity-set framework with tractable reformulations and conditional finite-sample/asymptotic guarantees.
The main claim is a conditional expressivity/memory separation with explicit hybrid constructions; learned-model experiments separately examine whether the theoretical advantage survives optimization.
The main contribution is an equilibrium impossibility and relaxed-existence analysis, with simulations supporting interpretation.
The contribution is a constrained-equilibrium existence theorem and a conditional convergence analysis for independent learning.
The contribution is a theoretical characterization and counterexample analysis of ordinal and dynamic stability in learning games.
The main contribution is a random-matrix analysis of finite-data denoiser and sampling-map consistency, with empirical tests of its predictions.
The principal contribution is a constructive architectural equivalence theorem and general controllability framework.
The main contribution is a stagewise theoretical account of attention learning, supported by mechanistic diagnostics.
The principal contribution is a conditional learning-theoretic framework for graph algorithms, with constructive and negative cases.
The main new knowledge is a controlled measurement, diagnostic, comparison, failure analysis or audit. A proposed remedy, model or tool can follow from that evidence.
Corpus context · 23 papers: median 6 main figures and 2 main tables; body span ≈ 9.04 pages; attached appendix span ≈ 14.91 pages (23 resolved boundaries). Separate source files: median Not resolved whole-file pages among 0 papers; these can repeat main text and are not appendix-only lengths. Extraction-assisted observations—not drafting targets.
Organize evidence around competing explanations and retain informative nulls or collapse regimes. Factor the intervention, distinguish uncertainty from checkpoint training versus evaluation resampling, and report the scope of unsuccessful or negative controls. Selected models, seeds and evaluation perturbations do not establish universal mechanism or causal law. These are lessons from the named examples, not frequency estimates for every member.
Editorial starting package: choose artifacts that test your actual claims. Quantities below are planning suggestions, not counts inferred from all category members or conference rules.
One compact view
Expose information access, sampling, intervention and matched controls.
When: Essential when teacher labels, dataset splits, task direction or model access determine the interpretation.
One clear entry point
Display the relationship, denominator and variability before a proposed explanation.
When: Show representative ranges rather than selected successes alone.
One destination per central explanatory claim
Distinguish the proposed mechanism from compute, label quality, dataset composition or evaluation artifacts.
When: Correlation and a fitted trend require different language from a controlled causal intervention.
A small aligned set with selection criteria
Make measured failures inspectable and include ambiguous/negative cases where relevant.
When: Qualitative cases illustrate a study; they do not replace prevalence estimates.
As needed by generalization or practicality claims
Repeat the protocol on different tasks/models/settings and report the actual cost.
When: Keep missing, unavailable or diverged results distinct from zero.
Full control matrices, data and prompt construction, thresholds, seed-level measurements, tuning and compute accounting, alternative metrics, all model/dataset results and known false-positive or selection effects.
Organize a diagnostic paper around separable components and preserve failed tuning settings alongside peak performance.
Move from descriptive diagnostics to targeted interventions, and label uncertainty by its actual source rather than using a generic multi-seed claim.
Define competing failure modes through observable cross-checks, then choose interventions that distinguish incentive effects from incidental representation drift.
Boundary: This route contains 23 primary editorial assignments in the 168-paper inspected subset (168 inventory records). It is not an official track, a universal writing standard or evidence that these choices caused acceptance.
See inline examples of these artifact styles · Compare every member in the measurement ledger
Controlled optimizer comparisons and update-sparsity diagnosis organize the contribution; no new optimizer is required.
Controlled behavioral-cloning learning dynamics and transfer findings organize the paper; synthetic tasks are instruments for that study.
The main argument diagnoses a behavior through controlled swaps and interventions; the curated benchmark enables this empirical test.
A failure-mode diagnosis and controlled gradient/proxy investigation organize the argument; filtering is the diagnostic's tested intervention.
A controlled taxonomy of training outcomes and mechanism-isolating probe experiments, rather than a new detector architecture, organize the contribution.
The central contribution is a controlled spectral investigation of domain adaptation; the software package and rewind interventions support that analysis.
The principal argument uses controlled architecture/order comparisons to clarify formulation-versus-implementation trade-offs.
The central argument diagnoses a generation-order trade-off through controlled comparisons before proposing a simplified training recipe.
The main argument evaluates generalization and training-mixture scaling of activation explanation across auditing tasks.
The paper studies emergent sequential-decision bias and compares interventions rather than introducing a new general learning architecture.
The central contribution links training examples to component emergence through attribution and controlled intervention analyses.
The central contribution is a mechanistic diagnosis supported by controlled interventions and a secondary detection validation.
The principal contribution is a controlled study of latent-action supervision choices and their task-dependent effects.
The contribution is an empirical investigation of pretrained policies, replay, forgetting metrics and recovery.
The primary contribution develops and empirically tests internal-flow indicators of response reliability.
The primary contribution studies a finite-state abstraction of reasoning and uses intervention experiments to test its utility.
Question-led controlled experiments on mixture, timing, retention and representation change drive the paper; the theoretical sketch motivates rather than replaces these tests.
The organizing question is why meta-learned data selection fails; formal gradient analysis motivates targeted batch-size and feature interventions, followed by controlled tests.
The paper adapts rare-event sampling into an end-to-end practical analysis and demonstrates estimation/error diagnostics on a tractable language model.
A formalized measurement framework supports controlled empirical studies of memorization capacity and membership scaling; the central observed claims remain dependent on architecture, training and reference models.
The primary contribution is an observational characterization of attainable post-training performance across compute and time; quantile fitting and evaluation selection support that empirical analysis.
The primary evidence is a mixed-methods observational study of practitioners and deployed systems, organized around research questions rather than a new model or benchmark leaderboard.
The contribution is an observational and controlled-prompt study of annotation behavior, using competing familiarity proxies and definition interventions rather than a new annotation model.
A new corpus, acquisition/annotation process, evaluation suite or measurement protocol is the primary reusable deliverable. A companion model demonstrates utility without replacing the need to validate the resource.
Corpus context · 17 papers: median 6 main figures and 3 main tables; body span ≈ 9.14 pages; attached appendix span ≈ 23.91 pages (17 resolved boundaries). Separate source files: median Not resolved whole-file pages among 0 papers; these can repeat main text and are not appendix-only lengths. Extraction-assisted observations—not drafting targets.
Expose the construct, label-production process and evaluation access before treating a score as capability. Turn the motivating taxonomy into testable axes, document oracle or simulator assistance, and connect failures back to those axes. Oracle-assisted human judgments, robot execution, process monitoring and outcome prediction are different evaluation contracts. These are lessons from the named examples, not frequency estimates for every member.
Editorial starting package: choose artifacts that test your actual claims. Quantities below are planning suggestions, not counts inferred from all category members or conference rules.
One compact resource table
Compare capabilities, annotation units, acquisition conditions and evaluation tasks, not only total size.
When: Use explicit partial or unavailable coverage rather than forcing binary superiority.
One process diagram with examples
Show raw inputs, transformations, human/machine verification and final annotations.
When: Distinguish generated labels, human corrections and measured ground truth.
A distribution, hierarchy or protocol matrix suited to the labels
Show subjects/scenes/classes, long tails, acquisition diversity and what is held out.
When: Avoid conflating frames, tracks, objects, rendered variants and independent subjects.
Focused tables with the same taxonomy used in the protocol
Demonstrate what the resource reveals and expose non-dominating models or hard subgroups.
When: A high average score alone does not validate label quality or benchmark relevance.
One verification analysis when annotation or metric validity is a claim
Document agreement, quality checks, adversarial controls, contamination checks or calibration.
When: Select the check that addresses this resource's actual failure risks; do not imply a single check proves complete validity.
Annotation instructions and interfaces, full templates where needed, inclusion/exclusion rules, split lists or generation settings, quality-control sampling, dataset statistics, baseline configuration, access conditions and detailed limitations.
Make the benchmark's construct unavoidable by design, then cross representation and integration choices while separating oracle and executable protocols.
Separate the construct, observation access and labeling mechanism before plotting scaling trends, and explain when aggregation changes the evaluated subset.
Boundary: This route contains 17 primary editorial assignments in the 168-paper inspected subset (168 inventory records). It is not an official track, a universal writing standard or evidence that these choices caused acceptance.
See inline examples of these artifact styles · Compare every member in the measurement ledger
The primary reusable deliverable is an evaluation protocol and operational measurement of answering versus benchmarking strength.
MADQA's annotated document collection, grounded evaluation and search diagnostics form the reusable resource; agent comparison demonstrates its measurement purpose.
The reusable benchmark and capability-oriented task annotations are the central deliverables; evaluations diagnose task and environment weaknesses.
The contribution is a protein fitness benchmark with realistic generalization splits; model comparisons characterize its difficulty rather than introduce a new model.
The paper explicitly centers TEDBench, with masked invariant autoencoding positioned as a reference method demonstrating the benchmark's utility.
Parameterized, solver-verifiable problem families and semantics-preserving interventions are the reusable evaluation contribution.
The reusable evaluation taxonomy, monitor scopes and measurement protocol anchor the scaling and intervention experiments.
A reusable value representation, intensity database and calibrated evaluation protocol are the central infrastructure supporting steering analyses.
The history-dependent manipulation benchmark and controlled memory-design comparison are the reusable contribution; model variants serve the construct investigation.
The principal artifact is a reproducible benchmarking framework and standardized evaluation resource.
The contribution introduces a controlled capability benchmark with validity and containment methodology.
A task-specific safety benchmark and its measurement protocol anchor the empirical contribution.
A controlled model collection and multilingual perturbation benchmark enable the diagnostic comparisons; their construction and measurement contract are the central reusable contribution.
The reusable contribution is an interactive causal benchmark with controlled observation pitfalls and multidimensional evaluation.
The primary reusable artifact is a large spatial-supervision dataset generated by an automated curation pipeline, evaluated both for annotation quality and downstream VLM utility.
The contribution is an executable conversational-agent evaluation environment with task generation and reliability diagnostics, not a new agent model.
The principal artifact is a pipeline and corpus of executable repair tasks, with cross-validation, benchmark construction and downstream training establishing data utility.
The central contribution changes execution, inference, memory, acquisition cost or resource scaling while preserving a stated quality or correctness contract.
Corpus context · 16 papers: median 5 main figures and 3 main tables; body span ≈ 8.98 pages; attached appendix span ≈ 15.91 pages (15 resolved boundaries). Separate source files: median Not resolved whole-file pages among 0 papers; these can repeat main text and are not appendix-only lengths. Extraction-assisted observations—not drafting targets.
Define the cost boundary and quality operating point as part of the result. Separate proxies from elapsed time, list excluded work, show workload regimes, and pair acceleration with the accuracy or optimality lost at that setting. A sub-second GPU region is not full user-visible latency; a pruned solver does not inherit an exact parent formulation's guarantee. These are lessons from the named examples, not frequency estimates for every member.
Editorial starting package: choose artifacts that test your actual claims. Quantities below are planning suggestions, not counts inferred from all category members or conference rules.
One focused breakdown or controlled comparison
Explain which stage is costly and why the proposed decision variable could help.
When: Use matched budgets when an oracle analysis motivates an approximate implementation.
One overview with needed stage or memory insets
Mark preprocessing, cache/reuse, training versus inference, hardware calls and synchronization.
When: The figure should expose added overhead as well as removed work.
One panel per materially different workload
Pair quality with latency, throughput, memory or acquisition effort at comparable operating points.
When: State hardware, precision, batch/sequence/image size and what the timer includes.
A table or stacked breakdown
Show whether an isolated saving survives the entire pipeline.
When: Do not substitute FLOPs, token count or number of steps for wall-clock measurements without qualification.
Curves only for the dimensions the system claims to scale
Measure view/frame count, problem size, cache refresh or device constraints and retain adverse regimes.
When: Changing hardware or input capacity is a different comparison from identical-workload acceleration.
As needed for the execution contract
Make reuse, selection, fallback, stopping and approximation explicit.
When: A claimed exact or distribution-preserving procedure needs a precise argument beyond similar output scores.
Hardware/software configuration, workload sizes, baseline settings, preprocessing and warm-up treatment, timing repeats, memory definitions, tuning budget, approximation sensitivity and failure regimes. Include overhead introduced by the method.
Attach every speed claim to its timed boundary and operating point, and expose unequal input information before comparing task accuracy.
Expose the full route from relaxation to executable output, and plot approximation quality alongside system speed instead of transferring exact guarantees to the accelerated pipeline.
Organize a serving paper around the actual scarce resource and workload regimes, with controlled baseline implementations and explicit deployment limits.
Boundary: This route contains 16 primary editorial assignments in the 168-paper inspected subset (168 inventory records). It is not an official track, a universal writing standard or evidence that these choices caused acceptance.
See inline examples of these artifact styles · Compare every member in the measurement ledger
The operational contribution is a low-overhead early-exit inference criterion; geometric observations and conditional bounds support the runtime intervention.
Batch-level token-budget scheduling and serving-kernel integration are primary; sparse confidence gating enables that systems objective.
The main contribution co-designs sampling, representation loops and precision around an explicitly timed molecular-docking operating point.
An exact-to-relaxed-to-projected optimization pipeline is evaluated through scalable runtime and solution-quality tradeoffs, supported by structural integrality results.
The central deliverable is a low-bit deployment method with calibration and runtime efficiency evidence.
The contribution scales an existing training parameterization through memory and kernel engineering.
The main method changes inference execution and measures the overhead of input-specific layer programs.
The contribution maps differential-equation computation to a physical dynamical-system design and evaluates accuracy and estimated hardware efficiency.
The argument joins trajectory construction, ordinary-engine orchestration and a training objective around a latency–accuracy trade-off.
Cache-compatible attention, training and streaming decoding are organized around practical inference efficiency rather than a standalone new benchmark.
The paper explicitly coordinates a multi-view federated detection pipeline with measured runtime and communication efficiency, alongside predictive effectiveness.
The main intervention makes multi-token attribution computationally tractable through aggregation and recursion; efficiency must be evaluated together with attribution quality and approximation limits.
The main contribution is a compiler and hardware-aware approximation pipeline for faster deployed activation functions.
The contribution co-designs fused GPU kernels and transport operators for reduced-memory, high-throughput entropic optimal transport.
The contribution is a GPU-aware sketch distribution and fused kernel for faster approximate linear algebra.
The contribution is a unified differentiable-programming library and implementation framework with systematic operator and performance comparisons.
The primary argument is a task-specific capability, physical measurement/reconstruction system or domain-grounded method whose validity depends on real acquisition, task assumptions and domain-appropriate evidence.
Corpus context · 8 papers: median 5.5 main figures and 3 main tables; body span ≈ 9.04 pages; attached appendix span ≈ 9.91 pages (8 resolved boundaries). Separate source files: median Not resolved whole-file pages among 0 papers; these can repeat main text and are not appendix-only lengths. Extraction-assisted observations—not drafting targets.
Write the domain's actual inputs, interventions and human work into the comparison contract. Show the workflow and access assumptions, assess domain-relevant outcomes, and preserve the failure or coverage cost of optimizing a preferred proxy. Chemical plausibility scores, medical task benchmarks or reviewed constraint graphs are not experimental or clinical deployment validation. These are lessons from the named examples, not frequency estimates for every member.
Editorial starting package: choose artifacts that test your actual claims. Quantities below are planning suggestions, not counts inferred from all category members or conference rules.
One setup or input/output view
Show sensors/modalities, calibration, measured versus generated quantities and what is available at use time.
When: Essential when hardware, multimodal cohorts or nonstandard labels determine the claim.
One overview with necessary physical or anatomical labels
Connect variables, modules and losses to the domain problem rather than only showing generic network boxes.
When: Keep measured inputs, latent physical quantities and predicted outputs visually distinct.
A compact grid or profile plot
Compare matched scenes/patients/objects with relevant maps, spectra, overlays or error views.
When: Label simulation, real data, pseudo-ground-truth and unavailable measurements; an appealing image is not independent validation.
One table per incompatible endpoint
Report the relevant physical/clinical/behavioral metric, units, uncertainty and evaluation split.
When: Separate image quality from geometry, survival discrimination, calibration or downstream behavior.
Controls chosen for the proposed mechanism
Test key physical assumptions, modality use or pipeline components; include an independent reference when feasible.
When: Do not label a model-derived pseudo-label as independent truth or infer clinical benefit from a technical benchmark.
As appropriate to the deployment claim
State calibration, capture, annotation, compute and failure requirements.
When: A successful laboratory demonstration does not establish unrestricted field or clinical reliability.
Acquisition/calibration details, specimen or cohort preparation, split and metric definitions, physical derivations, implementation settings, supplementary real/synthetic comparisons and failure cases. Keep pivotal ground-truth limitations in the main text.
For domain planning, pair improved multi-objective quality with coverage losses and disclose how learned or judged objective proxies are constructed.
Separate knowledge, procedure and control flow, then design ablations that change one integration choice at a time.
Boundary: This route contains 8 primary editorial assignments in the 168-paper inspected subset (168 inventory records). It is not an official track, a universal writing standard or evidence that these choices caused acceptance.
See inline examples of these artifact styles · Compare every member in the measurement ledger
The principal argument adapts world-model policy learning to sparse physical-field sensing with an explicit sensor/reconstruction contract.
The principal contribution addresses a specific video-editing deployment requirement, separating paired-video training from mask-free subtitle removal.
The contribution integrates planning, tools and aesthetic feedback for autonomous photo editing; the dataset and reward model support this workflow.
The primary argument applies structured expert procedures to domain reasoning through a controlled retrieval-and-generation workflow.
The primary writing burden connects multi-objective search to chemically relevant synthesis planning and its practical objective tradeoffs.
The main contribution adapts generative dynamical emulators to a practical Bayesian-filtering application across fluid and weather systems.
The method is organized around phenotype-conditioned connectome prediction and domain-specific interpretation, with architectural innovations serving that application.
The contribution is a black-box multimodal reliability-assessment pipeline combining object-focused prompts and an embedding statistic, evaluated as hallucination detection rather than base-model training.
The main contribution is a defensible stance or proposed change in research practice, not necessarily a new algorithm or benchmark. Official position-track obligations remain edition-specific.
Corpus context · 9 papers: median 2 main figures and 0 main tables; body span ≈ 9.04 pages; attached appendix span ≈ 5.92 pages (5 resolved boundaries). Separate source files: median Not resolved whole-file pages among 0 papers; these can repeat main text and are not appendix-only lengths. Extraction-assisted observations—not drafting targets.
Turn a thesis into an explicit evidence standard while stating what it does not prohibit. Define the conceptual distinction, connect supporting examples to evidence levels, and answer serious alternative views with concessions and a bounded response. These are writing lessons from Position Track papers; the edition-specific Alternative Views requirement belongs in official rules, and illustrative stress tests are not comprehensive safety assessments. These are lessons from the named examples, not frequency estimates for every member.
Editorial starting package: choose artifacts that test your actual claims. Quantities below are planning suggestions, not counts inferred from all category members or conference rules.
One compact map if relationships are otherwise difficult to follow
Show how premises support recommendations and who bears the benefits, costs and obligations.
When: Useful for multi-party proposals; a simple thesis does not need a decorative flowchart.
A focused comparison when precise distinctions matter
Compare assumptions, evidence, benefits, trade-offs and unresolved objections without caricaturing opposing views.
When: Use published facts or explicitly labeled reasoning; do not invent numerical support for a normative claim.
Selected cases with stated selection limits
Make the problem and implications inspectable, and distinguish illustrative cases from representative samples.
When: A counterexample can refute a universal claim but cannot establish how common a failure is.
A concise plan or table if several actors and steps are involved
Connect proposed changes to feasible actions, responsible actors, costs and observable success criteria.
When: Do not present proposed outcomes as measured results. A method leaderboard is not required merely to resemble a conventional ML paper.
Supporting cases, extended evidence, definitions, survey or search methods, and qualified implementation scenarios. Keep the central thesis and strongest alternatives in the main argument. Check the actual conference track's required sections; ICML 2026 Position Papers specifically require Alternative Views.
Turn a position into a testable agenda with explicit criteria, carefully scoped examples and fair concessions to alternative priorities.
Turn a methodological position into an explicit claim-to-evidence ladder, test plausible alternative explanations, and state what the thesis does not prohibit.
Boundary: This route contains 9 primary editorial assignments in the 168-paper inspected subset (168 inventory records). It is not an official track, a universal writing standard or evidence that these choices caused acceptance.
See inline examples of these artifact styles · Compare every member in the measurement ledger
Official Position Track paper advances a research-priority thesis through literature examples and counterarguments rather than presenting a new algorithm or benchmark.
Official Position Track paper whose contribution is an evidence-level thesis, research checklist and response to objections, supported by illustrative experiments.
The core contribution is a normative position and community call to action supported by literature and an affiliation analysis.
The contribution is a position about governance and evaluation of peer-review automation, supported by empirical demonstrations.
The central deliverable is a stakeholder-oriented position and research agenda for AI-assisted peer review.
The paper advances a normative community position supported by literature and cases, with explicit alternative views and recommendations.
The paper advances an institutional thesis and recommendations supported by conceptual distinctions, historical examples and explicit counterarguments.
An official Position-track paper arguing for a research and governance direction through synthesis, alternatives and a call to action.
An official Position-track argument for changing deepfake research priorities, supported by a literature landscape and structured alternative views.
