ICLR 2026 — the 223-paper writing atlas
Start with the distribution, then compare a paper like yours. This atlas measures the pinned final-paper PDFs behind all 223 oral writing studies. It complements—not replaces—their argument, visual-design and evidence analysis. A long appendix or many figures is not a standard to meet, and an oral-only corpus cannot explain what causes acceptance.
Choose your contribution type · See the inline visual gallery · Download the full measurement ledger (JSON)
223 / 223inspected PDFs / oral inventory records
7,236physical pages in those PDFs
6editorial contribution types
7separate files, excluded from totals
What the measured manuscripts look like
Observed extraction results, not a target length or figure quota. The middle 50% is the interval from the 25th to the 75th percentile. Missing boundaries are excluded from that span’s denominator, not replaced by zero. Caption counts are extraction-assisted and can miss nonstandard or image-only labels; inspect each paper’s notes.
| Measure | Median | Middle 50% | Full range | Denominator |
|---|---|---|---|---|
| Inspected PDF lengthphysical pages | 30 | 23–38 | 11–88 | 223 papers |
| Before references / appendixestimated page-equivalents | 10.17 | 9.88–10.32 | 8.51–11.11 | 223 papers |
| Attached appendix spanestimated page-equivalents | 15.89 | 9.89–22.88 | 0.21–72.49 | 218 papers |
| Main figurescaption labels | 5 | 3–6 | 0–14 | 223 papers |
| Main tablescaption labels | 3 | 1–4 | 0–9 | 223 papers |
| Appendix figurescaption labels | 6 | 2–10 | 0–53 | 223 papers |
| Appendix tablescaption labels | 4 | 2–9 | 0–68 | 223 papers |
| Separate source-file lengthwhole-file physical pages, possibly including repeated main text; not appendix-only length | 20 | 9.5–26.5 | 3–28 | 7 papers |
| Separate source-file figureswhole-file caption labels; may include overlapping main-text displays | 9 | 3.5–11.5 | 0–25 | 7 papers |
| Separate source-file tableswhole-file caption labels; may include overlapping main-text displays | 5 | 3–8.5 | 0–16 | 7 papers |
Compare like contributions
Each measured paper has one primary editorial type so the totals sum to 223. Secondary types preserve mixed contributions. These are classifications of the annotated argument, not official ICLR tracks. The table uses the same extraction method and shows medians; the category guides below explain what these artifacts should accomplish.
| Primary contribution | Papers | Main figures / tables | Body span ≈ pp. | Appendix span ≈ pp. |
|---|---|---|---|---|
| Architecture & representation design | 40 | 5 / 4 | 10.11 | 8.8939/40 attached boundaries; separate PDF median 18.5 pp. (2 papers) |
| Learning & optimization algorithms | 70 | 5 / 3 | 10.2 | 16.3669/70 attached boundaries; separate PDF median 3 pp. (1 papers) |
| Inference, algorithms & systems efficiency | 23 | 4 / 3 | 10.17 | 11.8921/23 attached boundaries; separate PDF median 26.5 pp. (2 papers) |
| Theory & formal foundations | 33 | 3 / 1 | 10.11 | 21.3932/33 attached boundaries; separate PDF median 10 pp. (1 papers) |
| Empirical analysis, diagnosis & auditing | 28 | 6 / 1 | 10.19 | 14.8728/28 attached boundaries; separate PDF median 20 pp. (1 papers) |
| Datasets, benchmarks & measurement resources | 29 | 5 / 3 | 10.26 | 21.929/29 attached boundaries; separate PDF median Not resolved pp. (0 papers) |
Search all 223 measurement records
Open a paper’s ledger for every detected top-level section, its estimated length and share of the PDF, and figure/table locations. Its existing study explains what those visuals and sections do.
Showing 223 of 223 records.
| Paper / category | PDF pages | Main F / T | Appendix F / T | Body / appendix span ≈ pp. | Notes |
|---|---|---|---|---|---|
| Latent Speech-Text TransformerArchitecture & representation design | 18 | 5 / 6 | 1 / 7 | 10.24 / 5.25 | No automatic flags |
| TROLL: Trust Regions Improve Reinforcement Learning for Large Language ModelsLearning & optimization algorithms | 40 | 5 / 1 | 13 / 4 | 10.1 / 24.36 | No automatic flags |
| On the Wasserstein Geodesic Principal Component Analysis of probability measuresTheory & formal foundations | 32 | 7 / 0 | 14 / 3 | 10.21 / 18.89 | No automatic flags |
| Probabilistic Kernel Function for Fast Angle TestingInference, algorithms & systems efficiency | 23 | 2 / 1 | 7 / 4 | 9.49 / 10.89 | No automatic flags |
| Multiplayer Nash Preference OptimizationLearning & optimization algorithms | 21 | 1 / 4 | 0 / 6 | 10.25 / 5.89 | No automatic flags |
| Shoot First, Ask Questions Later? Building Rational Agents that Explore and Act Like PeopleLearning & optimization algorithms | 46 | 5 / 1 | 15 / 8 | 10.11 / 28.8 | No automatic flags |
| GEPA: Reflective Prompt Evolution Can Outperform Reinforcement LearningLearning & optimization algorithms | 87 | 4 / 3 | 23 / 1 | 9.29 / 70.89 | No automatic flags |
| OpenThoughts: Data Recipes for Reasoning ModelsDatasets, benchmarks & measurement resources | 72 | 3 / 7 | 32 / 41 | 9.87 / 52.89 | No automatic flagsSource correction retained |
| Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics ModelingArchitecture & representation design | 49 | 4 / 3 | 13 / 10 | 10.46 / 28.89 | No automatic flags |
| Mastering Sparse CUDA Generation through Pretrained Models and Deep Reinforcement LearningLearning & optimization algorithms | 38 | 5 / 3 | 3 / 17 | 10.2 / 22.65 | No automatic flags |
| Universal Inverse Distillation for Matching Models with Real-Data Supervision (No GANs)Learning & optimization algorithms | 43 | 2 / 4 | 13 / 4 | 10.35 / 28.65 | No automatic flags |
| Partition Generative Modeling: Masked Modeling Without MasksArchitecture & representation design | 26 | 4 / 2 | 2 / 8 | 9.23 / 12.65 | No automatic flags |
| RedTeamCUA: Realistic Adversarial Testing of Computer-Use Agents in Hybrid Web-OS EnvironmentsDatasets, benchmarks & measurement resources | 46 | 3 / 3 | 7 / 10 | 10.74 / 28.89 | No automatic flags |
| RealPDEBench: A Benchmark for Complex Physical Systems with Real-World DataDatasets, benchmarks & measurement resources | 46 | 6 / 1 | 15 / 8 | 10.46 / 28.89 | No automatic flags |
| CauKer: Classification Time Series Foundation Models Can Be Pretrained on Synthetic DataDatasets, benchmarks & measurement resources | 29 | 7 / 2 | 6 / 11 | 10.23 / 14.74 | No automatic flags |
| The Shape of Adversarial Influence: Characterizing LLM Latent Spaces with Persistent HomologyEmpirical analysis, diagnosis & auditing | 47 | 10 / 3 | 53 / 6 | 10.28 / 31.16 | No automatic flags |
| mCLM: A Modular Chemical Language Model that Generates Functional and Makeable MoleculesArchitecture & representation design | 33 | 5 / 2 | 8 / 9 | 11.11 / 13.89 | No automatic flags |
| AdAEM: An Adaptively and Automated Extensible Measurement of LLMs' Value DifferenceDatasets, benchmarks & measurement resources | 49 | 9 / 1 | 12 / 17 | 10.62 / 27.89 | No automatic flags |
| Discount Model Search for Quality Diversity Optimization in High-Dimensional Measure SpacesLearning & optimization algorithms | 39 | 2 / 2 | 10 / 5 | 10.42 / 20.89 | No automatic flags |
| Modality-free Graph In-context AlignmentLearning & optimization algorithms | 27 | 4 / 5 | 1 / 8 | 10.21 / 11.89 | No automatic flags |
| Revela: Dense Retriever Learning via Language ModelingLearning & optimization algorithms | 24 | 5 / 2 | 2 / 17 | 10.11 / 9.89 | No automatic flags |
| Monocular Normal Estimation via Shading Sequence EstimationArchitecture & representation design | 13 | 6 / 6 | 0 / 0 | 10.11 / Unresolved | No attached appendix boundary detectedSeparate files: 28 pp.; 25 figures / 11 tables; excluded from main totals |
| Multimodal Aligned Semantic Knowledge for Unpaired Image-text MatchingArchitecture & representation design | 28 | 3 / 6 | 0 / 9 | 10.18 / 14.89 | No automatic flags |
| Exploring Synthesizable Chemical Space with Iterative Pathway RefinementsLearning & optimization algorithms | 23 | 6 / 5 | 4 / 4 | 9.82 / 9.9 | No automatic flags |
| Energy-Based Transformers are Scalable Learners and ThinkersArchitecture & representation design | 44 | 8 / 5 | 9 / 5 | 10.45 / 25.7 | No automatic flags |
| Hubble: a Model Suite to Advance the Study of LLM MemorizationDatasets, benchmarks & measurement resources | 62 | 3 / 1 | 20 / 14 | 10.46 / 38.87 | No automatic flags |
| Softmax Transformers are Turing-CompleteTheory & formal foundations | 21 | 1 / 1 | 0 / 2 | 10.21 / 7.62 | No automatic flags |
| It's All Just Vectorization: einx, a Universal Notation for Tensor OperationsInference, algorithms & systems efficiency | 20 | 0 / 2 | 0 / 2 | 10.11 / 7.89 | No automatic flags |
| PateGAIL++: Utility Optimized Private Trajectory Generation with Imitation LearningLearning & optimization algorithms | 16 | 2 / 7 | 1 / 3 | 10.11 / 2.89 | No automatic flags |
| Coupling Experts and Routers in Mixture-of-Experts via an Auxiliary LossLearning & optimization algorithms | 21 | 7 / 1 | 3 / 2 | 10.11 / 6.54 | No automatic flags |
| Exploratory Causal Inference in SAEnceLearning & optimization algorithms | 48 | 6 / 0 | 31 / 2 | 11.11 / 32.9 | No automatic flags |
| TileLang: Bridge Programmability and Performance in Modern Neural KernelsInference, algorithms & systems efficiency | 36 | 10 / 0 | 9 / 19 | 10.36 / 21.54 | No automatic flagsSource correction retained |
| Characterizing the Discrete Geometry of ReLU NetworksTheory & formal foundations | 27 | 7 / 1 | 7 / 2 | 10.11 / 11.89 | No automatic flags |
| Gaussian certified unlearning in high dimensions: A hypothesis testing approachTheory & formal foundations | 38 | 5 / 1 | 9 / 0 | 10.11 / 23.89 | No automatic flags |
| InfoNCE Induces Gaussian DistributionTheory & formal foundations | 22 | 4 / 2 | 7 / 0 | 10.42 / 8.27 | No automatic flagsSource correction retained |
| MomaGraph: State-Aware Unified Scene Graphs with Vision-Language Models for Embodied Task PlanningArchitecture & representation design | 26 | 5 / 3 | 14 / 3 | 10.34 / 10.89 | No automatic flags |
| Veritas: Generalizable Deepfake Detection via Pattern-Aware ReasoningLearning & optimization algorithms | 58 | 6 / 8 | 34 / 11 | 10.21 / 40.89 | No automatic flags |
| FRABench and UFEval: Unified Fine-grained Evaluation with Task and Aspect GeneralizationDatasets, benchmarks & measurement resources | 72 | 5 / 7 | 18 / 68 | 10.18 / 56.24 | Figure numbering discontinuity; see source noteSource correction retained |
| Omni-Reward: Towards Generalist Omni-Modal Reward Modeling with Free-Form PreferencesLearning & optimization algorithms | 50 | 3 / 3 | 14 / 17 | 11.11 / 28.89 | No automatic flagsSource correction retained |
| Half-order Fine-Tuning for Diffusion Model: A Recursive Likelihood Ratio OptimizerLearning & optimization algorithms | 27 | 6 / 5 | 5 / 5 | 9.2 / 12.89 | No automatic flags |
| A Scalable Distributed Framework for Multimodal GigaVoxel Image RegistrationInference, algorithms & systems efficiency | 20 | 8 / 1 | 0 / 0 | 10.18 / Unresolved | No attached appendix boundary detectedSource correction retainedSeparate files: 26 pp.; 12 figures / 4 tables; excluded from main totals |
| Seeing Through the Brain: New Insights from Decoding Visual Stimuli with fMRIArchitecture & representation design | 25 | 2 / 6 | 5 / 3 | 10.32 / 8.89 | No automatic flags |
| FIRE: Frobenius-Isometry Reinitialization for Balancing the Stability–Plasticity TradeoffLearning & optimization algorithms | 33 | 5 / 0 | 3 / 10 | 10.26 / 17.89 | No automatic flags |
| Stable Video Infinity: Infinite-Length Video Generation with Error RecyclingLearning & optimization algorithms | 27 | 9 / 4 | 9 / 8 | 10.2 / 13.89 | No automatic flags |
| Temporal superposition and feature geometry of RNNs under memory demandsTheory & formal foundations | 42 | 5 / 0 | 12 / 1 | 10.3 / 25.9 | No automatic flags |
| Visual Planning: Let's Think Only with ImagesLearning & optimization algorithms | 34 | 6 / 2 | 10 / 11 | 9.44 / 18.89 | No automatic flags |
| EditBench: Evaluating LLM Abilities to Perform Real-World Instructed Code EditsDatasets, benchmarks & measurement resources | 27 | 5 / 3 | 10 / 5 | 8.51 / 13.89 | No automatic flags |
| UALM: Unified Audio Language Model for Understanding, Generation and ReasoningArchitecture & representation design | 23 | 6 / 4 | 1 / 4 | 10.34 / 5.89 | No automatic flags |
| Overcoming Joint Intractability with Lossless Hierarchical Speculative DecodingInference, algorithms & systems efficiency | 34 | 2 / 4 | 1 / 4 | 10.21 / 21.4 | No automatic flags |
| Fast training of accurate physics-informed neural networks without gradient descentLearning & optimization algorithms | 60 | 7 / 2 | 23 / 32 | 10.29 / 41.89 | No automatic flags |
| RefineStat: Efficient Exploration for Probabilistic Program SynthesisInference, algorithms & systems efficiency | 24 | 2 / 6 | 0 / 3 | 10.17 / 9.89 | No automatic flags |
| How Learning Rate Decay Wastes Your Best Data in Curriculum-Based LLM PretrainingEmpirical analysis, diagnosis & auditing | 32 | 6 / 2 | 3 / 10 | 10.21 / 14.89 | No automatic flags |
| DiffusionNFT: Online Diffusion Reinforcement with Forward ProcessLearning & optimization algorithms | 22 | 10 / 1 | 3 / 1 | 9.58 / 8.89 | No automatic flags |
| High-dimensional Analysis of Synthetic Data SelectionTheory & formal foundations | 52 | 1 / 3 | 1 / 10 | 10.16 / 36.89 | No automatic flags |
| Navigating the Latent Space Dynamics of Neural ModelsEmpirical analysis, diagnosis & auditing | 29 | 5 / 0 | 13 / 4 | 10.67 / 12.89 | No automatic flags |
| LLMs Get Lost In Multi-Turn ConversationEmpirical analysis, diagnosis & auditing | 41 | 5 / 1 | 5 / 9 | 10.74 / 24.89 | No automatic flags |
| Sequences of Logits Reveal the Low Rank Structure of Language ModelsEmpirical analysis, diagnosis & auditing | 37 | 6 / 0 | 6 / 3 | 10.3 / 22.35 | No automatic flags |
| MedAgentGym: A Scalable Agentic Training Environment for Code-Centric Reasoning in Biomedical Data ScienceDatasets, benchmarks & measurement resources | 43 | 7 / 5 | 10 / 8 | 10.11 / 24.89 | No automatic flags |
| BIRD-INTERACT: Re-imagining Text-to-SQL Evaluation via Lens of Dynamic InteractionsDatasets, benchmarks & measurement resources | 49 | 7 / 3 | 19 / 8 | 10.73 / 29.62 | No automatic flags |
| Pareto-Conditioned Diffusion Models for Offline Multi-Objective OptimizationLearning & optimization algorithms | 33 | 3 / 2 | 6 / 30 | 10.39 / 17.89 | No automatic flags |
| Enhancing Generative Auto-bidding with Offline Reward Evaluation and Policy SearchLearning & optimization algorithms | 32 | 5 / 5 | 8 / 8 | 10.11 / 15.89 | No automatic flags |
| Overparametrization bends the landscape: BBP transitions at initialization in simple Neural NetworksTheory & formal foundations | 26 | 4 / 0 | 6 / 0 | 10.25 / 13.59 | No automatic flags |
| Learning to See Before Seeing: Demystifying LLM Visual Priors from Language Pre-trainingEmpirical analysis, diagnosis & auditing | 31 | 8 / 3 | 6 / 3 | 10.39 / 13.65 | No automatic flags |
| Quotient-Space Diffusion ModelsTheory & formal foundations | 43 | 3 / 3 | 0 / 0 | 9.21 / 27.89 | No automatic flags |
| Overthinking Reduction with Decoupled Rewards and Curriculum Data SchedulingLearning & optimization algorithms | 30 | 4 / 2 | 11 / 9 | 9.86 / 16.36 | No automatic flags |
| Verifying Chain-of-Thought Reasoning via Its Computational GraphEmpirical analysis, diagnosis & auditing | 28 | 4 / 4 | 4 / 15 | 9.88 / 12.89 | No automatic flags |
| Differentially Private Domain DiscoveryTheory & formal foundations | 27 | 3 / 0 | 10 / 1 | 9.29 / 16.34 | No automatic flags |
| Instilling an Active Mind in Avatars via Cognitive SimulationArchitecture & representation design | 33 | 5 / 4 | 8 / 2 | 10.11 / 15.89 | No automatic flags |
| Decentralized Attention Fails Centralized Signals: Rethinking Transformers for Medical Time SeriesArchitecture & representation design | 21 | 5 / 5 | 0 / 7 | 10.11 / 5.89 | No automatic flags |
| Pre-training under infinite computeEmpirical analysis, diagnosis & auditing | 41 | 9 / 0 | 16 / 12 | 10.11 / 19.84 | No automatic flags |
| ParaRNN: Unlocking Parallel Training of Nonlinear RNNs for Large Language ModelsInference, algorithms & systems efficiency | 32 | 3 / 2 | 10 / 4 | 10.54 / 17.89 | No automatic flagsSource correction retained |
| World-In-World: World Models in a Closed-Loop WorldDatasets, benchmarks & measurement resources | 40 | 7 / 6 | 11 / 5 | 10.14 / 22.83 | No automatic flags |
| Mean Flow Policy with Instantaneous Velocity Constraint for One-step Action GenerationLearning & optimization algorithms | 23 | 5 / 3 | 3 / 3 | 9.83 / 9.89 | No automatic flags |
| Huxley-Gödel Machine: Human-Level Coding Agent Development by an Approximation of the Optimal Self-Improving MachineLearning & optimization algorithms | 31 | 1 / 3 | 0 / 1 | 10.65 / 16.89 | No automatic flags |
| PhyWorldBench: A Comprehensive Evaluation of Physical Realism in Text-to-Video ModelsDatasets, benchmarks & measurement resources | 35 | 6 / 4 | 15 / 13 | 9.21 / 23.28 | No automatic flags |
| On The Surprising Effectiveness of a Single Global Merging in Decentralized LearningEmpirical analysis, diagnosis & auditing | 36 | 2 / 1 | 6 / 0 | 10.11 / 18.29 | No automatic flags |
| Depth Anything 3: Recovering the Visual Space from Any ViewsArchitecture & representation design | 25 | 6 / 6 | 3 / 8 | 10.11 / 8.89 | No automatic flagsSource correction retained |
| Watch your steps: Dormant Adversarial Behaviors that Activate upon LLM FinetuningEmpirical analysis, diagnosis & auditing | 40 | 4 / 9 | 31 / 11 | 10.11 / 24.4 | No automatic flags |
| MetaEmbed: Scaling Multimodal Retrieval at Test-Time with Flexible Late InteractionArchitecture & representation design | 20 | 4 / 3 | 0 / 4 | 9.47 / 3.89 | No automatic flags |
| Extending Sequence Length is Not All You Need: Effective Integration of Multimodal Signals for Gene Expression PredictionLearning & optimization algorithms | 22 | 5 / 3 | 1 / 13 | 9.87 / 9.15 | No automatic flags |
| Improving Diffusion Models for Class-imbalanced Training Data via Capacity ManipulationLearning & optimization algorithms | 40 | 4 / 7 | 12 / 15 | 10.2 / 23.88 | No automatic flags |
| Intrinsic Entropy of Context Length Scaling in LLMsTheory & formal foundations | 36 | 7 / 0 | 11 / 2 | 9.52 / 20.89 | No automatic flags |
| Non-Asymptotic Analysis of (Sticky) Track-and-StopTheory & formal foundations | 31 | 0 / 0 | 0 / 0 | 10.34 / 18.89 | No automatic flagsSource correction retained |
| GLASS Flows: Efficient Inference for Reward Alignment of Flow and Diffusion ModelsInference, algorithms & systems efficiency | 35 | 3 / 2 | 11 / 3 | 10.11 / 20.89 | No automatic flags |
| MotionStream: Real-Time Video Generation with Interactive Motion ControlsInference, algorithms & systems efficiency | 26 | 8 / 4 | 6 / 4 | 10.11 / 7.89 | No automatic flags |
| Agent Data Protocol: Unifying Datasets for Diverse, Effective Fine-tuning of LLM AgentsDatasets, benchmarks & measurement resources | 21 | 2 / 8 | 2 / 3 | 10.27 / 6.19 | No automatic flags |
| LLM Fingerprinting via Semantically Conditioned WatermarksLearning & optimization algorithms | 36 | 2 / 3 | 10 / 6 | 10.11 / 19.89 | No automatic flags |
| Triple-BERT: Do We Really Need MARL for Order Dispatch on Ride-Sharing Platforms?Architecture & representation design | 29 | 3 / 2 | 3 / 8 | 10.52 / 12.89 | No automatic flags |
| Hallucination Begins Where Saliency DropsInference, algorithms & systems efficiency | 22 | 4 / 3 | 10 / 3 | 9.11 / 8.34 | No automatic flags |
| To Infinity and Beyond: Tool-Use Unlocks Length Generalization in State Space ModelsTheory & formal foundations | 27 | 3 / 1 | 7 / 1 | 10.11 / 12.73 | No automatic flags |
| Reducing Belief Deviation in Reinforcement Learning for Active Reasoning of LLM AgentsLearning & optimization algorithms | 35 | 6 / 3 | 9 / 5 | 10.21 / 20.48 | Table numbering discontinuity; see source noteSource correction retained |
| Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time ComputeArchitecture & representation design | 54 | 11 / 3 | 21 / 11 | 10.85 / 36.87 | No automatic flagsSource correction retained |
| The Spacetime of Diffusion Models: An Information Geometry PerspectiveTheory & formal foundations | 25 | 7 / 1 | 1 / 0 | 9.36 / 13.62 | No automatic flags |
| Learning to Segment for Vehicle Routing ProblemsLearning & optimization algorithms | 37 | 7 / 5 | 5 / 11 | 10.37 / 22.89 | No automatic flagsSource correction retained |
| Mixture-of-Experts Can Surpass Dense LLMs Under Strictly Equal ResourceEmpirical analysis, diagnosis & auditing | 22 | 3 / 2 | 2 / 15 | 10.11 / 7.18 | No automatic flagsSeparate files: 20 pp.; 5 figures / 16 tables; excluded from main totals |
| LoongRL: Reinforcement Learning for Advanced Reasoning over Long ContextsLearning & optimization algorithms | 20 | 4 / 6 | 5 / 4 | 9.63 / 6.89 | No automatic flags |
| SANA-Video: Efficient Video Generation with Block Linear Diffusion TransformerArchitecture & representation design | 25 | 6 / 6 | 9 / 3 | 10.11 / 9.89 | No automatic flagsSeparate files: 9 pp.; 9 figures / 2 tables; excluded from main totals |
| Learning with Dual-level Noisy Correspondence for Multi-modal Entity AlignmentLearning & optimization algorithms | 32 | 5 / 3 | 10 / 15 | 9.54 / 19.9 | No automatic flags |
| Optimistic Task Inference for Behavior Foundation ModelsLearning & optimization algorithms | 34 | 6 / 0 | 14 / 2 | 9.35 / 21.89 | No automatic flags |
| Online Learning and Equilibrium Computation with Ranking FeedbackTheory & formal foundations | 46 | 2 / 1 | 9 / 0 | 10.59 / 31.9 | No automatic flags |
| Radiometrically Consistent Gaussian Surfels for Inverse RenderingLearning & optimization algorithms | 33 | 6 / 2 | 14 / 10 | 9.46 / 19.89 | No automatic flags |
| From Markov to Laplace: How Mamba In-Context Learns Markov ChainsTheory & formal foundations | 31 | 3 / 1 | 10 / 8 | 9.47 / 17.89 | No automatic flags |
| AutoEP: LLMs-Driven Automation of Hyperparameter Evolution for Metaheuristic AlgorithmsLearning & optimization algorithms | 19 | 6 / 4 | 9 / 4 | 10.2 / 5.84 | No automatic flags |
| P-GenRM: Personalized Generative Reward Model with Test-time User-based ScalingLearning & optimization algorithms | 29 | 4 / 5 | 11 / 10 | 10.18 / 15.89 | No automatic flags |
| Locality-aware Parallel Decoding for Efficient Autoregressive Image GenerationInference, algorithms & systems efficiency | 24 | 9 / 3 | 5 / 4 | 10.11 / 7.89 | No automatic flags |
| Global Resolution: Optimal Multi-Draft Speculative Sampling via Convex OptimizationInference, algorithms & systems efficiency | 34 | 1 / 2 | 4 / 3 | 10.11 / 20.89 | No automatic flags |
| Semi-Supervised Preference Optimization with Limited FeedbackLearning & optimization algorithms | 30 | 3 / 5 | 0 / 14 | 10.19 / 15.89 | No automatic flags |
| A Representer Theorem for Hawkes Processes via Penalized Least Squares MinimizationTheory & formal foundations | 24 | 0 / 4 | 3 / 4 | 10.23 / 10.89 | No automatic flags |
| Diffusion Language Model Knows the Answer Before It DecodesInference, algorithms & systems efficiency | 19 | 3 / 4 | 2 / 3 | 10.11 / 4.89 | No automatic flags |
| EigenBench: A Comparative Behavioral Measure of Value AlignmentDatasets, benchmarks & measurement resources | 29 | 3 / 4 | 7 / 14 | 10.2 / 16.9 | No automatic flags |
| In-Place Test-Time TrainingLearning & optimization algorithms | 21 | 4 / 3 | 0 / 5 | 10.43 / 5.89 | No automatic flags |
| OpenApps: Simulating Environment Variations to Measure UI Agent ReliabilityDatasets, benchmarks & measurement resources | 34 | 6 / 0 | 4 / 13 | 10.11 / 17.89 | No automatic flags |
| Generating metamers of human scene understandingEmpirical analysis, diagnosis & auditing | 26 | 5 / 0 | 11 / 0 | 10.39 / 10.89 | No automatic flags |
| Addressing divergent representations from causal interventions on neural networksEmpirical analysis, diagnosis & auditing | 27 | 3 / 0 | 6 / 0 | 10.33 / 12.89 | No automatic flags |
| Token-Importance Guided Direct Preference OptimizationLearning & optimization algorithms | 23 | 4 / 2 | 2 / 13 | 10.48 / 9.6 | No automatic flags |
| Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language for InterpretabilityArchitecture & representation design | 29 | 5 / 2 | 7 / 8 | 10.52 / 14.89 | No automatic flags |
| EditVerse: Unifying Image and Video Editing and Generation with In-Context LearningArchitecture & representation design | 22 | 8 / 5 | 2 / 6 | 10.16 / 5.89 | No automatic flags |
| Compositional Diffusion with Guided search for Long-Horizon PlanningInference, algorithms & systems efficiency | 15 | 7 / 5 | 0 / 0 | 10.46 / Unresolved | No attached appendix boundary detectedSeparate files: 27 pp.; 11 figures / 6 tables; excluded from main totals |
| Distributional Equivalence in Linear Non-Gaussian Latent-Variable Cyclic Causal Models: Characterization and LearningTheory & formal foundations | 48 | 3 / 0 | 5 / 5 | 10.34 / 29.87 | No automatic flags |
| Uncover Underlying Correspondence for Robust Multi-view ClusteringLearning & optimization algorithms | 20 | 3 / 2 | 5 / 2 | 9.86 / 6.89 | No automatic flags |
| AgentGym-RL: An Open-Source Framework to Train LLM Agents for Long-Horizon Decision Making via Multi-Turn RLDatasets, benchmarks & measurement resources | 43 | 8 / 4 | 14 / 4 | 10.46 / 20.89 | No automatic flags |
| Generative Human Geometry DistributionArchitecture & representation design | 20 | 9 / 5 | 12 / 0 | 9.19 / 5.89 | No automatic flags |
| TabStruct: Measuring Structural Fidelity of Tabular DataDatasets, benchmarks & measurement resources | 63 | 4 / 2 | 6 / 64 | 10.86 / 44.87 | No automatic flags |
| Reliable Weak-to-Strong Monitoring of LLM AgentsEmpirical analysis, diagnosis & auditing | 56 | 5 / 0 | 18 / 3 | 9.71 / 41.89 | No automatic flags |
| Reasoning with Sampling: Your Base Model is Smarter Than You ThinkInference, algorithms & systems efficiency | 19 | 5 / 2 | 4 / 3 | 9.81 / 5.89 | No automatic flags |
| LLM DNA: Tracing Model Evolution via Functional RepresentationsEmpirical analysis, diagnosis & auditing | 28 | 6 / 4 | 4 / 5 | 10.49 / 14.87 | No automatic flags |
| Information Shapes Koopman RepresentationLearning & optimization algorithms | 50 | 6 / 2 | 16 / 10 | 10.28 / 35.79 | No automatic flags |
| TD-JEPA: Latent-predictive Representations for Zero-Shot Reinforcement LearningLearning & optimization algorithms | 34 | 4 / 1 | 9 / 5 | 10.11 / 19.89 | No automatic flags |
| Steering the Herd: A Framework for LLM-based Control of Social LearningTheory & formal foundations | 38 | 2 / 0 | 3 / 1 | 10.19 / 23.9 | No automatic flags |
| DTO-KD: Dynamic Trade-off Optimization for Effective Knowledge DistillationLearning & optimization algorithms | 13 | 4 / 5 | 0 / 0 | 10.11 / Unresolved | No attached appendix boundary detectedSeparate files: 3 pp.; 0 figures / 0 tables; excluded from main totals |
| RAIN-Merging: A Gradient-Free Method to Enhance Instruction Following in Large Reasoning Models with Preserved Thinking FormatLearning & optimization algorithms | 41 | 6 / 6 | 7 / 10 | 10.25 / 24.89 | No automatic flagsSource correction retained |
| SafeDPO: A Simple Approach to Direct Preference Optimization with Enhanced SafetyLearning & optimization algorithms | 40 | 3 / 3 | 1 / 26 | 10.11 / 27.9 | No automatic flags |
| Beyond Prompt-Induced Lies: Investigating LLM Deception on Benign PromptsEmpirical analysis, diagnosis & auditing | 29 | 7 / 0 | 16 / 5 | 10.49 / 14.87 | No automatic flags |
| Speculative Actions: A Lossless Framework for Faster AI AgentsInference, algorithms & systems efficiency | 24 | 6 / 1 | 3 / 1 | 10.17 / 11.89 | No automatic flagsSource correction retained |
| Train-before-Test Harmonizes Language Model RankingsEmpirical analysis, diagnosis & auditing | 33 | 8 / 2 | 7 / 6 | 9.56 / 15.89 | No automatic flags |
| One for Two: A Unified Framework for Imbalanced Graph Classification via Dynamic Balanced PrototypeLearning & optimization algorithms | 38 | 6 / 5 | 10 / 19 | 10.43 / 21.87 | No automatic flags |
| WAVE: Learning Unified & Versatile Audio-Visual Embeddings with Multimodal LLMArchitecture & representation design | 17 | 1 / 7 | 1 / 3 | 10.11 / 2.7 | No automatic flags |
| Actions Speak Louder than Prompts: A Large-Scale Study of LLMs for Graph InferenceEmpirical analysis, diagnosis & auditing | 23 | 4 / 3 | 0 / 9 | 9.42 / 9.89 | No automatic flags |
| In-the-Flow Agentic System Optimization for Effective Planning and Tool UseLearning & optimization algorithms | 47 | 9 / 4 | 3 / 0 | 10.23 / 31.89 | No automatic flags |
| Q-RAG: Long Context Multi‑Step Retrieval via Value‑Based Embedder TrainingLearning & optimization algorithms | 26 | 3 / 4 | 3 / 4 | 10.11 / 12.89 | No automatic flags |
| ThinKV: Thought-Adaptive KV Cache Compression for Efficient Reasoning ModelsInference, algorithms & systems efficiency | 35 | 11 / 5 | 7 / 10 | 10.2 / 20.31 | No automatic flagsSource correction retained |
| DCFold: Efficient Protein Structure Generation with Single Forward PassLearning & optimization algorithms | 18 | 7 / 6 | 0 / 3 | 10.34 / 4.24 | No automatic flags |
| LongWriter-Zero: Mastering Ultra-Long Text Generation via Reinforcement LearningLearning & optimization algorithms | 26 | 6 / 1 | 0 / 4 | 10.62 / 12.37 | No automatic flags |
| MC-Search: Evaluating and Enhancing Multimodal Agentic Search with Structured Long Reasoning ChainsDatasets, benchmarks & measurement resources | 35 | 5 / 4 | 5 / 10 | 10.11 / 21.9 | No automatic flags |
| Mamba-3: Improved Sequence Modeling using State Space PrinciplesArchitecture & representation design | 30 | 2 / 4 | 4 / 8 | 10.19 / 12.86 | No automatic flags |
| WSM: Decay-Free Learning Rate Schedule via Checkpoint Merging for LLM Pre-trainingLearning & optimization algorithms | 29 | 5 / 5 | 6 / 5 | 9.72 / 11.89 | No automatic flags |
| FlashVID: Efficient Video Large Language Models via Training-free Tree-based Spatiotemporal Token MergingInference, algorithms & systems efficiency | 30 | 4 / 6 | 6 / 8 | 10.11 / 13.9 | No automatic flags |
| Is it Thinking or Cheating? Detecting Implicit Reward Hacking by Measuring Reasoning EffortEmpirical analysis, diagnosis & auditing | 26 | 14 / 0 | 17 / 2 | 9.79 / 12.89 | No automatic flags |
| Multi-Domain Riemannian Graph Gluing for Building Graph Foundation ModelsArchitecture & representation design | 41 | 6 / 1 | 2 / 21 | 10.11 / 22.89 | No automatic flags |
| FALCON: Few-step Accurate Likelihoods for Continuous FlowsLearning & optimization algorithms | 25 | 7 / 5 | 10 / 1 | 11.11 / 9.89 | No automatic flags |
| Pinet: Optimizing hard-constrained neural networks with orthogonal projection layersArchitecture & representation design | 47 | 4 / 0 | 18 / 6 | 9.82 / 28.89 | No automatic flags |
| SAFETY-GUIDED FLOW (SGF): A UNIFIED FRAMEWORK FOR NEGATIVE GUIDANCE IN SAFE GENERATIONInference, algorithms & systems efficiency | 28 | 4 / 4 | 8 / 2 | 10.57 / 13.89 | No automatic flags |
| Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality AssessmentEmpirical analysis, diagnosis & auditing | 19 | 6 / 5 | 3 / 6 | 9.11 / 4.89 | No automatic flags |
| On the Generalization Capacities of MLLMs for Spatial IntelligenceArchitecture & representation design | 15 | 7 / 5 | 0 / 1 | 9.11 / 2.89 | No automatic flagsSource correction retained |
| Task-free Adaptive Meta Black-box OptimizationLearning & optimization algorithms | 31 | 5 / 2 | 11 / 6 | 9.44 / 17.89 | No automatic flags |
| How Do Transformers Learn to Associate Tokens: Gradient Leading Terms Bring Mechanistic InterpretabilityTheory & formal foundations | 39 | 7 / 1 | 1 / 2 | 10.62 / 24.89 | No automatic flags |
| Compactness and Consistency: A Conjoint Framework for Deep Graph ClusteringLearning & optimization algorithms | 24 | 5 / 7 | 2 / 3 | 9.82 / 7.89 | No automatic flags |
| Taming Momentum: Rethinking Optimizer States Through Low-Rank ApproximationLearning & optimization algorithms | 24 | 3 / 3 | 0 / 1 | 9.47 / 11.64 | No automatic flags |
| CounselBench: A Large-Scale Expert Evaluation and Adversarial Benchmarking of Large Language Models in Mental Health Question AnsweringDatasets, benchmarks & measurement resources | 37 | 5 / 4 | 3 / 14 | 10.47 / 18.26 | No automatic flags |
| Difficult Examples Hurt Unsupervised Contrastive Learning: A Theoretical PerspectiveTheory & formal foundations | 31 | 4 / 6 | 2 / 5 | 10.11 / 17.89 | No automatic flagsSource correction retained |
| Spherical Watermark: Encryption-Free, Lossless Watermarking for Diffusion ModelsInference, algorithms & systems efficiency | 33 | 6 / 5 | 11 / 12 | 10.45 / 18.31 | No automatic flags |
| Visual symbolic mechanisms: Emergent symbol processing in Vision Language ModelsEmpirical analysis, diagnosis & auditing | 32 | 7 / 0 | 40 / 1 | 10.32 / 18.89 | No automatic flags |
| CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at ScaleDatasets, benchmarks & measurement resources | 46 | 8 / 2 | 17 / 7 | 10.82 / 30.89 | No automatic flagsSource correction retained |
| FlashWorld: High-quality 3D Scene Generation within SecondsLearning & optimization algorithms | 23 | 7 / 3 | 6 / 0 | 9.59 / 8.29 | No automatic flags |
| Plug-and-Play Compositionality for Boosting Continual Learning with Foundation ModelsLearning & optimization algorithms | 34 | 4 / 2 | 10 / 11 | 10.4 / 17.89 | No automatic flags |
| MrRoPE: Mixed-radix Rotary Position EmbeddingArchitecture & representation design | 16 | 6 / 3 | 2 / 2 | 9.44 / 4.56 | No automatic flags |
| BioX-Bridge: Model Bridging for Unsupervised Cross-Modal Knowledge Transfer across BiosignalsArchitecture & representation design | 27 | 4 / 3 | 10 / 9 | 9.6 / 13.89 | No automatic flags |
| Efficient Resource-Constrained Training of Transformers via Subspace OptimizationLearning & optimization algorithms | 21 | 8 / 0 | 4 / 4 | 9.36 / 8.89 | No automatic flags |
| The Art of Scaling Reinforcement Learning Compute for LLMsEmpirical analysis, diagnosis & auditing | 30 | 6 / 0 | 20 / 1 | 9.8 / 15.89 | No automatic flags |
| The Polar Express: Optimal Matrix Sign Methods and their Application to the Muon AlgorithmLearning & optimization algorithms | 38 | 6 / 0 | 12 / 0 | 10.23 / 22.89 | No automatic flags |
| DepthLM: Metric Depth from Vision Language ModelsArchitecture & representation design | 15 | 8 / 3 | 2 / 3 | 9.65 / 1.89 | No automatic flags |
| Exchangeability of GNN Representations with Applications to Graph RetrievalTheory & formal foundations | 66 | 4 / 1 | 17 / 3 | 10.28 / 48.89 | Figure numbering discontinuity; see source note; Table numbering discontinuity; see source noteSource correction retained |
| Benchmarking Empirical Privacy Protection for Adaptations of Large Language ModelsDatasets, benchmarks & measurement resources | 38 | 6 / 3 | 8 / 37 | 10.25 / 19.89 | No automatic flags |
| From movement to cognitive maps: recurrent neural networks reveal how locomotor development shapes hippocampal spatial codingEmpirical analysis, diagnosis & auditing | 23 | 5 / 0 | 9 / 4 | 10.1 / 9.9 | No automatic flags |
| Gaia2: Benchmarking LLM Agents on Dynamic and Asynchronous EnvironmentsDatasets, benchmarks & measurement resources | 32 | 10 / 3 | 10 / 4 | 10.26 / 17.89 | No automatic flags |
| SimuHome: A Temporal- and Environment-Aware Benchmark for Smart Home LLM AgentsDatasets, benchmarks & measurement resources | 28 | 5 / 3 | 1 / 11 | 10.28 / 14.89 | No automatic flags |
| Planner Aware Path Learning in Diffusion Language Models TrainingLearning & optimization algorithms | 46 | 4 / 4 | 2 / 3 | 11.11 / 30.89 | No automatic flags |
| Latent Fourier TransformArchitecture & representation design | 32 | 5 / 1 | 8 / 11 | 10.24 / 14.89 | No automatic flags |
| ScaleCUA: Scaling Open-Source Computer Use Agents with Cross-Platform DataDatasets, benchmarks & measurement resources | 60 | 6 / 4 | 26 / 16 | 10.55 / 42.89 | No automatic flags |
| Transformers are Inherently SuccinctTheory & formal foundations | 21 | 0 / 0 | 0 / 0 | 10.19 / 7.89 | No automatic flags |
| Let Features Decide Their Own Solvers: Hybrid Feature Caching for Diffusion TransformersInference, algorithms & systems efficiency | 24 | 8 / 5 | 7 / 0 | 10.66 / 9.89 | No automatic flags |
| Optimal Sparsity of Mixture-of-Experts Language Models for Reasoning TasksEmpirical analysis, diagnosis & auditing | 25 | 8 / 1 | 17 / 3 | 10.21 / 9.45 | No automatic flags |
| HATSolver: Learning Gröbner Bases with Hierarchical Attention TransformersArchitecture & representation design | 21 | 2 / 3 | 4 / 9 | 10.16 / 8.43 | No automatic flags |
| On the Reasoning Abilities of Masked Diffusion Language ModelsTheory & formal foundations | 36 | 3 / 0 | 0 / 0 | 10.3 / 21.89 | No automatic flags |
| Rodrigues Network for Learning Robot ActionsArchitecture & representation design | 23 | 5 / 3 | 1 / 9 | 9.43 / 9.89 | No automatic flags |
| TRACE: Your Diffusion Model is Secretly an Instance Edge DetectorLearning & optimization algorithms | 39 | 11 / 7 | 10 / 9 | 10.28 / 21.9 | No automatic flags |
| AnyUp: Universal Feature UpsamplingArchitecture & representation design | 21 | 6 / 9 | 6 / 2 | 10.11 / 6.89 | No automatic flags |
| WAFT: Warping-Alone Field Transforms for Optical FlowArchitecture & representation design | 14 | 4 / 6 | 0 / 1 | 9.27 / 0.21 | No automatic flags |
| Fast Escape, Slow Convergence: Learning Dynamics of Phase Retrieval under Power-Law DataTheory & formal foundations | 43 | 3 / 0 | 2 / 1 | 10.11 / 29.82 | No automatic flags |
| WebDevJudge: Evaluating (M)LLMs as Critiques for Web Development QualityDatasets, benchmarks & measurement resources | 33 | 4 / 7 | 7 / 7 | 10.16 / 17.89 | No automatic flags |
| Generative Universal Verifier as Multimodal Meta-ReasonerLearning & optimization algorithms | 33 | 7 / 3 | 3 / 4 | 10.37 / 17.89 | No automatic flags |
| Non-Convex Federated Optimization under Cost-Aware Client SelectionTheory & formal foundations | 43 | 2 / 1 | 10 / 4 | 10.11 / 29.9 | No automatic flags |
| p-less Sampling: A Robust Hyperparameter-Free Approach for LLM DecodingInference, algorithms & systems efficiency | 44 | 4 / 4 | 15 / 11 | 10.27 / 30.89 | No automatic flags |
| Through the Lens of Contrast: Self-Improving Visual Reasoning in VLMsLearning & optimization algorithms | 19 | 6 / 4 | 3 / 0 | 9.33 / 3.42 | No automatic flags |
| MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-based Memory AgentLearning & optimization algorithms | 29 | 9 / 4 | 1 / 7 | 11.11 / 12.89 | No automatic flags |
| Text-to-3D by Stitching a Multi-view Reconstruction Network to a Video GeneratorArchitecture & representation design | 31 | 6 / 5 | 10 / 2 | 10.11 / 12.89 | No automatic flags |
| Vid-LLM: A Compact Video-based 3D Multimodal LLM with Reconstruction–Reasoning SynergyArchitecture & representation design | 20 | 5 / 6 | 2 / 3 | 10.31 / 3.89 | No automatic flags |
| EmotionThinker: Prosody-Aware Reinforcement Learning for Explainable Speech Emotion ReasoningLearning & optimization algorithms | 26 | 4 / 7 | 5 / 1 | 9.2 / 12.82 | No automatic flags |
| Every Language Model Has a Forgery-Resistant SignatureTheory & formal foundations | 17 | 7 / 2 | 4 / 0 | 9.37 / 4.63 | No automatic flagsSource correction retained |
| The Coverage Principle: How Pre-Training Enables Post-TrainingTheory & formal foundations | 69 | 2 / 0 | 0 / 1 | 10.2 / 50.81 | No automatic flags |
| VibeVoice: Expressive Podcast Generation with Next-Token DiffusionArchitecture & representation design | 22 | 3 / 3 | 3 / 8 | 9.16 / 8.67 | No automatic flags |
| TTSDS2: Resources and Benchmark for Evaluating Human-Quality Text to Speech SystemsDatasets, benchmarks & measurement resources | 20 | 3 / 5 | 5 / 3 | 10.11 / 5.89 | No automatic flags |
| What's In My Human Feedback? Learning Interpretable Descriptions of Preference DataEmpirical analysis, diagnosis & auditing | 36 | 3 / 1 | 6 / 17 | 10.27 / 21.89 | No automatic flags |
| AstaBench: Rigorous Benchmarking of AI Agents with a Scientific Research SuiteDatasets, benchmarks & measurement resources | 88 | 2 / 3 | 8 / 20 | 11.11 / 72.49 | No automatic flags |
| Cross-Domain Lossy Compression via Rate- and Classification-Constrained Optimal TransportTheory & formal foundations | 50 | 6 / 0 | 13 / 15 | 9.17 / 35.84 | No automatic flags |
| Causal Structure Learning in Hawkes Processes with Complex Latent Confounder NetworksTheory & formal foundations | 33 | 5 / 0 | 5 / 8 | 9.87 / 19.9 | No automatic flags |
| Neon: Negative Extrapolation From Self-Training Improves Image GenerationLearning & optimization algorithms | 57 | 10 / 0 | 23 / 2 | 10.1 / 41.9 | No automatic flagsSource correction retained |
| Exploratory Diffusion Model for Unsupervised Reinforcement LearningLearning & optimization algorithms | 31 | 4 / 1 | 5 / 5 | 10.46 / 14.89 | No automatic flags |
| How Reliable is Language Model Micro-Benchmarking?Empirical analysis, diagnosis & auditing | 24 | 6 / 1 | 12 / 4 | 10.24 / 10.89 | No automatic flags |
| Why DPO is a Misspecified Estimator and How to Fix ItTheory & formal foundations | 11 | 3 / 2 | 0 / 0 | 9.18 / Unresolved | No attached appendix boundary detectedSeparate files: 10 pp.; 2 figures / 5 tables; excluded from main totals |
| Conformal Robustness Control: A New Strategy for Robust DecisionLearning & optimization algorithms | 29 | 5 / 1 | 2 / 5 | 10.23 / 15.89 | No automatic flags |
| Differentiable Model Predictive Control on the GPUInference, algorithms & systems efficiency | 19 | 8 / 0 | 6 / 6 | 9.21 / 6.89 | No automatic flags |
| True Self-Supervised Novel View Synthesis is TransferableArchitecture & representation design | 17 | 1 / 3 | 2 / 6 | 9.56 / 4.7 | No automatic flags |
| SWINGARENA: Adversarial Programming Arena for Long-context GitHub Issue SolvingDatasets, benchmarks & measurement resources | 29 | 3 / 3 | 4 / 3 | 10.11 / 15.73 | No automatic flags |
| Quantitative Bounds for Length Generalization in TransformersTheory & formal foundations | 39 | 4 / 0 | 0 / 0 | 9.54 / 26.89 | No automatic flags |
| Hyperparameter Trajectory Inference with Conditional Lagrangian Optimal TransportLearning & optimization algorithms | 30 | 3 / 7 | 2 / 8 | 10.33 / 15.89 | No automatic flagsSource correction retained |
| NextStep-1: Toward Autoregressive Image Generation with Continuous Tokens at ScaleArchitecture & representation design | 23 | 4 / 6 | 5 / 3 | 10.11 / 6.89 | No automatic flags |
| Structured Flow Autoencoders: Learning Structured Probabilistic Representations with Flow MatchingArchitecture & representation design | 22 | 4 / 4 | 6 / 0 | 10.11 / 8.89 | No automatic flags |
| InfoTok: Adaptive Discrete Video Tokenizer via Information-Theoretic CompressionArchitecture & representation design | 23 | 4 / 3 | 2 / 3 | 10.29 / 9.9 | No automatic flags |
| Invisible Safety Threat: Malicious Finetuning for LLM via SteganographyEmpirical analysis, diagnosis & auditing | 30 | 6 / 0 | 3 / 9 | 10.11 / 15.89 | No automatic flags |
| Why Low-Precision Transformer Training Fails: An Analysis on Flash AttentionEmpirical analysis, diagnosis & auditing | 24 | 7 / 0 | 6 / 0 | 10.17 / 10.89 | No automatic flagsSource correction retained |
| Common Corpus: The Largest Collection of Ethical Data for LLM Pre-TrainingDatasets, benchmarks & measurement resources | 34 | 4 / 2 | 1 / 12 | 10.31 / 15.89 | No automatic flags |
| Scaling Laws and Spectra of Shallow Neural Networks in the Feature Learning RegimeTheory & formal foundations | 42 | 3 / 0 | 5 / 0 | 10.11 / 27.29 | No automatic flags |
No matching papers. Reset or use a shorter query.
Measurement method
Section titled “Measurement method”The inventory is the 2026-09-19 official oral-program snapshot. Each measurement records the source URL and physical page count of the PDF in its existing writing study. No new reading-completion status is inferred from extraction.
| Question | What is measured | What it does not mean |
|---|---|---|
| How many figures or tables? | Unique printed caption labels, separately for main text and attached appendix; continued captions are deduplicated. | Not the number of plotted panels, photographs, curves, diagrams, algorithms or theorem boxes. A captioned prompt box can be a figure. |
| How long is the paper? | Physical pages of the canonical conference PDF. | Not the anonymous submission page count or a claim of rule compliance. |
| How long is a section? | Distance from its detected top-level heading to the next, using physical page and vertical position; rounded to 0.01 page-equivalents. | Not text-only length, occupied-ink area or an exact word count. Floats, whitespace and omitted subsection boundaries remain within the span. |
| What fraction is the section? | Its estimated span divided by the entire canonical PDF length. | Not the share of a permitted submission budget; references and front matter remain in the denominator. |
| Where does the main text end? | The first reference or attached-appendix boundary. | Statements before references can be included; this is a layout envelope, not an official main-content-page calculation. |
| How long is the appendix? | From its identified opening, including an explicit appendix contents/overview page where corrected, to the canonical PDF end. | Not just prose, proof length or the sum of every linked supplement. Missing boundaries remain unresolved rather than zero. |
| Which style is most common? | Contribution-type membership is editorially coded from all existing full-text annotations; primary types are mutually exclusive and secondary types capture overlap. | A universal conference taxonomy or frequency counts of every visual style. The gallery is purpose-selected, not prevalence-ranked. |
Extraction uses Poppler word coordinates, caption patterns and top-level headings. Source-bound corrections handle split captions, unusual capitalization, genuine numbering gaps, shared figure/table numbering, and appendix navigation before section A. A diverse 12-paper count audit, plus targeted flagged-caption checks, is retained with the source data. All 223 PDFs were processed; every caption and section boundary has not been independently hand-counted. Even an unflagged record can contain an extraction error. The underlying paper studies’ prior full-text reading is a separate evidence layer.
Separate PDFs remain separately inspectable. In particular, the MoE equal-resource and SANA-Video linked files overlap older manuscript material; adding them would inflate corpus totals. Raw PDFs, full extracted text and inspection screenshots stay in the ignored local cache. The public ledger contains small structural metadata, not redistributed manuscripts.
How to use the numbers responsibly
Section titled “How to use the numbers responsibly”- Choose the contribution type that matches your main claim, not the model acronym.
- Use its distribution to understand variation, then draft a claim-to-evidence plan.
- Open two or three comparable papers’ section ledgers to inspect different space allocations.
- Choose a visual form because it answers a reader’s question—not to match a figure count.
- Apply the official rules for the exact submission phase, which take precedence over observed final-paper layouts.
The source data, classification rubric, corrections and extractor are versioned. Future editions should use their own inventory, source URLs and corrections; do not silently blend conference years or carry numerical targets forward.