Deep Residual Learning for Image Recognition
Official Best Paper · SourceShortcut connections let a network learn changes relative to its input, helping optimize deeper networks.
TOP CONFERENCE PAPERS
Top Conference Papers
Explore sourced AI research by conference, year, and recognition. Official awards include evidence; EyeCode Pick is a separate editorial selection.
CONFERENCE DIRECTORY
Choose a conference
PAPER LIBRARY
CVPR · 2026
Attention is All you Need
The Transformer uses attention to model sequences without recurrent or convolutional networks.
LOAM: Lidar Odometry and Mapping in Real-time
Separates fast motion estimation from slower mapping to process LiDAR measurements in real time.
Indiscernible Object Counting in Underwater Scenes
EyeCode Pick · editorial selectionStudies counting hard-to-distinguish objects in underwater scenes and introduces the IOCfish5K dataset.
Transformers are Inherently Succinct
Official Outstanding Paper · SourceStudies how compactly Transformers represent formal languages. The proofs compare their representation size with automata and temporal logic, while exposing the computational difficulty of model verification.
LLMs Get Lost In Multi-Turn Conversation
Official Outstanding Paper · SourceExamines failures when instructions arrive over multiple conversation turns. Simulated conversations separate task aptitude from reliability and trace how premature assumptions affect later responses.
The Polar Express: Optimal Matrix Sign Methods and their Application to the Muon Algorithm
Official Honorable Mention · SourceComputes the matrix polar decomposition used by Muon. Minimax-designed polynomial iterations use GPU-friendly matrix multiplications, with analysis of approximation error and finite-precision behavior.
Safety Alignment Should be Made More Than Just a Few Tokens Deep
Official Outstanding Paper · SourceInvestigates safety alignment concentrated in the first output tokens. Case studies and regularized fine-tuning objectives explore deeper alignment and its persistence under subsequent fine-tuning.
Learning Dynamics of LLM Finetuning
Official Outstanding Paper · SourceDecomposes prediction changes during language-model fine-tuning. The framework connects interactions between responses with repetition, hallucination, and decreasing preferred-response probability during prolonged DPO training.
AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models
Official Outstanding Paper · SourceUpdates language-model knowledge while limiting interference with retained facts. Parameter changes are projected into a null space associated with preserved knowledge, supported by analysis and sequential-editing experiments.
Data Shapley in One Training Run
Official Honorable Mention · SourceAttributes training data contributions to a particular model run. It accumulates Shapley values of gradient updates rather than repeatedly retraining models on different data subsets.
SAM 2: Segment Anything in Images and Videos
Official Honorable Mention · SourceExtends promptable segmentation from images to video. A Transformer with streaming memory works with an interactive data engine to maintain prompted object masks across frames.
Faster Cascades via Speculative Decoding
Official Honorable Mention · SourceCombines model cascades with speculative decoding. Deferral rules are implemented through speculative execution and parallel verification to study inference cost–quality trade-offs.
Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)
Official Best Paper · SourceIntroduces Infinity-Chat to study diversity in open-ended responses. The analysis examines repetition within models, similarity across models, and differences in human preferences that automated judges may miss.
Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free
Official Best Paper · SourceStudies gating in softmax attention. Head-specific sigmoid gates after scaled dot-product attention provide a controlled way to examine nonlinearity, sparsity, training stability, and attention sinks.
1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities
Official Best Paper · SourceStudies network depth in self-supervised goal-conditioned reinforcement learning. Locomotion and manipulation experiments examine how deep residual architectures and training scale affect goal-reaching behavior.
Why Diffusion Models Don’t Memorize: The Role of Implicit Dynamical Regularization in Training
Official Best Paper · SourceSeparates the onset of useful diffusion generation from later training-data memorization. Experiments and a random-features model analyze how training dynamics create a window of generalization before memorization.
Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
Official Best Paper Runner-up · SourceCompares base models and models trained with verifiable rewards on reasoning tasks. Pass@k and solution-distribution analyses distinguish improved sampling efficiency from the emergence of new reasoning patterns.
Optimal Mistake Bounds for Transductive Online Learning
Official Best Paper Runner-up · SourceStudies the value of seeing an unlabeled input sequence before online prediction. Matching bounds relate mistakes to Littlestone dimension and clarify the gap between transductive and standard online learning.
Superposition Yields Robust Neural Scaling
Official Best Paper Runner-up · SourceUses a toy model of overlapping features to investigate scaling laws. Geometric analysis and language-model measurements connect representation dimension, superposition strength, and prediction loss.
Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction
Official Best Paper · SourceModels image generation as predicting the next resolution scale. Coarse-to-fine generation replaces raster-order token prediction and supports studies of scaling and image editing.
Stochastic Taylor Derivative Estimator: Efficient amortization for arbitrary differential operators
Official Best Paper · SourceAddresses the cost of high-order differential operators in neural networks. Taylor-mode automatic differentiation and randomized tensor contractions estimate operators for applications such as physics-informed neural networks.
Not All Tokens Are What You Need for Pretraining
Official Best Paper Runner-up · SourceQuestions whether every pretraining token deserves equal weight. Selective Language Modeling scores tokens with a reference model and focuses training loss on tokens aligned with the target distribution.
Guiding a Diffusion Model with a Bad Version of Itself
Official Best Paper Runner-up · SourceStudies control of image quality separately from diversity. Autoguidance steers diffusion sampling using a smaller or less-trained version of the same model rather than an unconditional model.
The PRISM Alignment Dataset: What Participatory, Representative and Individualised Human Feedback Reveals About the Subjective and Multicultural Alignment of Large Language Models
Official Best Paper · SourceConnects participant demographics and stated preferences with feedback on language-model conversations. The dataset enables analysis of representation, cultural variation, and disagreement in alignment data.
Conformal Prediction as Bayesian Quadrature
Official Outstanding Paper · SourceRevisits conformal prediction from a Bayesian perspective on deployment risk. Bayesian quadrature represents plausible test-loss distributions beyond the information conveyed by frequentist guarantees alone.
Roll the dice & look before you leap: Going beyond the creative limits of next-token prediction
Official Outstanding Paper · SourceBuilds minimal algorithmic tasks for testing planning and diverse generation. Comparisons of next-token learning, multi-token methods, and input noise provide a controlled testbed for creative behavior.
CollabLLM: From Passive Responders to Active Collaborators
Official Outstanding Paper · SourceTrains models for collaboration across conversation turns rather than only the next response. Simulated interactions estimate longer-term rewards for clarifying user intent and supporting task completion.
Score Matching with Missing Data
Official Outstanding Paper · SourceExtends score matching to observations missing arbitrary subsets of coordinates. Importance-weighted and variational estimators enable distribution learning without fully observed data.
Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusions
Official Outstanding Paper · SourceAnalyzes the trade-off between difficult training-time infilling and flexible inference-time token order in masked diffusion. Theory and experiments show how adaptive decoding can avoid difficult subproblems.
The Value of Prediction in Identifying the Worst-Off
Official Outstanding Paper · SourceStudies prediction in policies prioritizing the worst-off. Mathematical models and a long-term unemployment case study compare better predictions with expanding administrative capacity to provide assistance.
The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models
Official Outstanding Paper · SourceTests whether arbitrary token order helps reasoning in diffusion language models. JustGRPO uses left-to-right RL rollouts to investigate avoidance of uncertain decisions while retaining parallel inference decoding.
High-Accuracy Sampling for Diffusion Models and Log-Concave Distributions
Official Outstanding Paper · SourceDevelops sampling algorithms using score or gradient queries for diffusion and log-concave distributions. Complexity analysis connects target accuracy with score error and assumptions about the data distribution.
The Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probes
Official Honorable Mention · SourceExamines coding models trained against deception detectors. A taxonomy distinguishes representation drift from detector-evading output strategies, with analysis of KL regularization and detector penalties.
Motion Attribution for Video Generation
Official Honorable Mention · SourceInvestigates which training clips influence generated motion. Gradient-based attribution uses motion-weighted masks to separate temporal dynamics from static appearance and guide fine-tuning data selection.
How much can language models memorize?
Official Honorable Mention · SourceSeparates memorization of particular data from generalization about its generating process. An information-based measure and capacity experiments study data size, memorization, and membership inference.
A Random Matrix Theory Perspective on the Consistency of Diffusion Models
Official Honorable Mention · SourceInvestigates why diffusion models trained on separate data splits can generate similar outputs from the same noise. Random-matrix analysis isolates shared Gaussian statistics and finite-data variability as a baseline for nonlinear models.
To Grok Grokking: Provable Grokking in Ridge Regression
Official Honorable Mention · SourceProves delayed generalization in ridge regression. Analysis of gradient descent with weight decay connects the delay to training hyperparameters and examines how tuning can amplify or remove grokking.
Efficiently Reconstructing Dynamic Scenes One D4RT at a Time
Best Paper · CVPR 2026 · SourceCombines depth, spatiotemporal correspondence, and camera estimation from dynamic video. A unified point-query interface replaces separate task decoders in a feed-forward approach to 4D scene reconstruction.
NitroGen: An Open Foundation Model for Generalist Gaming Agents
Best Paper Honorable Mention · CVPR 2026 · SourceStudies vision–action agents across games. Actions extracted from gameplay video support behavior cloning and a multi-game evaluation environment for examining transfer to unseen games.
SAM 3D: 3Dfy Anything in Images
Best Paper Honorable Mention · CVPR 2026 · SourceReconstructs object geometry, texture, and layout from a single image. Human- and model-assisted annotation combines synthetic pretraining with real-image alignment for cluttered and occluded scenes.
VGGT: Visual Geometry Grounded Transformer
Best Paper · CVPR 2025 · SourceUses one feed-forward network to infer cameras, depth, point maps, and tracks from one or many images. The shared architecture studies jointly learned 3D geometry instead of separate task-specific solvers.
MegaSaM: Accurate, Fast and Robust Structure and Motion from Casual Dynamic Videos
Best Paper Honorable Mention · CVPR 2025 · SourceEstimates cameras and depth from casual monocular videos with dynamic scenes and unconstrained camera motion. Training and inference changes to deep visual SLAM address violations of static-scene assumptions.
Navigation World Models
Best Paper Honorable Mention · CVPR 2025 · SourcePredicts future visual observations from history and navigation actions. A conditional diffusion Transformer simulates and evaluates trajectories, allowing planning constraints beyond a fixed navigation policy.
Molmo and PixMo: Open Weights and Open Data for State-of-the-Art Vision-Language Models
Best Paper Honorable Mention · CVPR 2025 · SourceStudies open-data, open-weight vision–language models. Human-collected captions, question answering, and pointing data provide a basis for examining model choices and training pipelines without distilling proprietary VLM outputs.
3D Student Splatting and Scooping
Best Paper Honorable Mention · CVPR 2025 · SourceRevisits mixture representations in 3D Gaussian Splatting. Student distributions with positive and negative densities, together with a sampling-based optimizer, explore scene representations beyond conventional Gaussian components.
Generating Physically Stable and Buildable Brick Structures from Text
Official Best Paper · SourceGenerates interlocking brick structures from text. An autoregressive model uses validity checks and physics-aware rollback to constrain designs for stability and assembly.
Spatially-Varying Autofocus
Official Honorable Mention · SourceStudies spatially selective focusing at different image locations. A Lohmann lens and phase-only spatial light modulator combine with contrast and disparity cues to iteratively adjust the focus map.
RayZer: A Self-supervised Large View Synthesis Model
Official Honorable Mention · SourceLearns 3D scene reconstruction from images without known cameras or 3D supervision. Disentangled camera and scene representations use ray structure and view reconstruction to learn from 2D images.
Passive Ultra-Wideband Single-Photon Imaging
Official Best Paper · SourceReconstructs changing light intensity from photon timestamps without source synchronization. Flux-estimation theory and a Fourier-domain algorithm analyze scene motion and illumination over widely different timescales.
Adding Conditional Control to Text-to-Image Diffusion Models
Official Best Paper · SourceAdds spatial conditioning to pretrained text-to-image diffusion models. A control branch with zero-initialized convolutions accepts edges, depth, or poses while preserving the original backbone.
Segment Anything
Official Honorable Mention · SourceDefines promptable image segmentation. A model, interactive data engine, and large mask dataset provide a foundation for segmentation across new images and tasks.
Minimalist Vision with Freeform Pixels
Official Best Paper · SourceDesigns task-specific cameras with freeform pixel shapes. Modeling hardware as a network’s first layer learns optical masks for low-measurement sensing with limited identifying visual detail.
Concept Arithmetics for Circumventing Concept Inhibition in Diffusion Models
Official Honorable Mention · SourceExamines limitations of concept inhibition in text-to-image diffusion. Analysis of compositional inference shows why combining concepts matters when evaluating model safety mechanisms.
Rasterized Edge Gradients: Handling Discontinuities Differentiably
Official Honorable Mention · SourceAddresses gradients at visibility discontinuities in differentiable rasterization. Micro-edges approximate the continuous process behind discrete pixels, enabling geometry optimization without altering the forward rendering.
Heat Kernel Textures: the Geodesic Gaussians That Do Not Splat
Official Best Paper · SourceRepresents textures directly on triangle meshes with anisotropic heat kernels. Surface-aware optimization and density control replace UV unwrapping when fitting existing textures or multi-view images.
LSRM: High-Fidelity Object-Centric Reconstruction via Scaled Context Windows
Official Honorable Mention · SourceScales context windows for object-centric 3D reconstruction using sparse attention. Coarse-to-fine prediction, geometry-aware routing, and sequence parallelism focus computation on informative regions for geometry and appearance recovery.
Poppy: Polarization-Based Plug-and-Play Guidance for Enhancing Monocular Normal Estimation
Official Honorable Mention · SourceRefines surface normals from frozen RGB backbones using polarization measurements. Input/output offsets and reflectance decomposition are optimized through differentiable rendering to add physical constraints at inference time.
OmniRetarget: Interaction-Preserving Data Generation for Humanoid Whole-Body Loco-Manipulation and Scene Interaction
Official Best Paper · SourceRetargets human motion to humanoids while preserving body–object–terrain relationships. An interaction mesh and kinematic constraints generate and augment motion data for loco-manipulation.
Marginalizing and Conditioning Gaussians onto Linear Approximations of Smooth Manifolds with Applications in Robotics
Official Best Paper · SourceDerives Gaussian marginalization and conditioning on general linear manifolds. Linearization extends the method to smooth manifolds, with covariance-estimation applications in SLAM and constrained optimization.
MAC-VO: Metrics-aware Covariance for Learning-based Stereo Visual Odometry
Official Best Paper · SourceIntroduces metric-aware uncertainty to stereo visual odometry. Learned covariance guides keypoint selection and residual weighting, connecting image-matching reliability with pose-graph optimization.
LT-Mem: Volatility-Aware Spatio-Temporal Memory for Lifelong Scene Understanding
Official Best Paper · SourceMaintains object identity and history across revisits to changing environments. Volatility-aware updates to Live, Delta, and Meta memories support temporal scene queries from aligned multi-session observations.
Interdigitated Electrodes for Selective Stimulation of Skeletal Muscle Actuators in Biosyncretic Robots
Official Best Paper · SourceStudies selective control of skeletal-muscle actuators in biosyncretic robots. Optimized thin-film interdigitated electrodes target different muscle regions to support coordinated actuation.
Learning a Unified Policy for Position and Force Control in Legged Loco-Manipulation
Official Best Paper · SourceLearns a unified position-and-force policy for legged robots. Simulated reinforcement learning uses state history to estimate forces and compensate motion, supporting contact-rich manipulation without force sensors.
Fabrica: Dual-Arm Assembly of General Multi-Part Objects via Integrated Planning and Learning
Official Best Paper · SourceCombines hierarchical planning and learned control for dual-arm assembly. Assembly sequences, grasps, and motion plans connect with local contact policies across multi-part object geometries.
PoliFormer: Scaling On-Policy RL with Transformers Results in Masterful Navigators
Official Outstanding Paper · SourceLearns indoor navigation from RGB using large-scale simulated reinforcement learning. A vision encoder and causal Transformer combine perception with history for studying transfer to real robots.
One Model to Drift Them All: Physics-Informed Conditional Diffusion Model for Driving at the Limits
Official Outstanding Paper · SourceControls vehicles near handling limits using conditional diffusion. A physics-informed dynamics-parameter distribution is conditioned on online measurements and integrated into model predictive control.
FlashSAC: Fast and Stable Off-Policy Reinforcement Learning for High-Dimensional Robot Control
Official Outstanding Paper · SourceStudies stable off-policy learning for high-dimensional robot control. Bounds on weight, feature, and gradient norms combine with parallel simulation and replay to limit critic-error amplification.
FEAST: A Flexible Mealtime-Assistance System Towards In-the-Wild Personalization
Official Outstanding Paper · SourcePersonalizes robotic mealtime assistance in homes. Modular hardware, multiple interaction methods, and parameterized behavior trees adapt feeding, drinking, and mouth-wiping to individual needs.
Native and Compact Structured Latents for 3D Generation
Best Student Paper · CVPR 2026 · SourceThe paper proposes O-Voxel, a sparse voxel representation holding both geometry and PBR material, and an SC-VAE with 16x compression so a textured 1024 asset needs about 9.6K tokens. Flow-matching models totaling about 4B parameters generate shape and material from an image, and the authors report gains over baselines in reconstruction and generation experiments.
ChordEdit: One-Step Low-Energy Transport for Image Editing
Best Student Paper Honorable Mention · CVPR 2026 · SourceChordEdit performs text-guided image editing in a single step for one-step image generators, without training or inversion, using a low-energy control field derived from dynamic optimal transport. On PIE-bench the full version runs in 0.38 seconds.
No verified papers match these filters. Choose another conference, year, or selection.