FuncGenFoil is a new generative model that designs and edits high-fidelity airfoils as continuous functions, outperforming prior methods in accuracy and diversity while enabling flexible, resolution-free shape generation.

We introduce a novel inference-time steering methodology called Reinforcing Cognitive Experts (RICE), designed to improve reasoning depth and efficiency without additional training or complex heuristics.

We propose a scalable 4D dynamic reconstruction model trained only on real-world monocular RGB videos.

We propose a novel transfer learning framework for regression where outputs are probability distributions residing in the Wasserstein space.

A watermark forgery attack

We identify a smoothness-generalization dilemma in message passing that limits GNN universality across varying homophily and propose the Inceptive Graph Neural Network (IGNN), a universal framework to address the dilemma.

This paper explore the essence of message passing on heterophilic graphs from the perspective of model and design a method with the guidance of theoretical findings.

E-MoFlow unifies egomotion and optical flow estimation through implicit regularizations under a self-supervised paradigm.

A dataset and an algorithm for multi-modal object tracking.

We reveal how return-coverage affects the performance of conditional sequence modeling policies in offline RL and propose an algorithm achieving new state-of-the-art results on D4RL.

This work introduces a novel framework that learns physical sample residuals instead of direct mappings for PDE solving, significantly improving neural operator generalization in data-limited scenarios.

CoIDO is a lightweight dual-objective data selection framework that jointly optimizes importance and diversity, achieving 98.2% of full-data performance for multimodal instruction tuning using only 20% of the data.

A selective watermarking method for LLMs via a lightweight network trained on multi-objectives

we propose Continuous Subspace Optimization for Continual Learning (CoSO) to fine-tune the pre-trained models in a series of subspaces rather than a single one.

RayFusion is a superior camera-only collaborative perception framework based on ray occupancy information fusion.

We introduce MERIT, the first multilingual dataset for interleaved multi-condition semantic retrieval, revealing current methods' limitations and proposing new method to enhance retrieval performance.

The paper provides a novel approach to ICL example selection/ordering through the extension of DPO.

Our approach improves Vision Transformer Dense Representations via Self-Distillation

We propose OmniGaze, a semi-supervised learning framework, which utilizes large-scale unlabeled data and reward-driven pseudo-labeling strategy to effectively generalize gaze estimation in the wild.

We propose a model-agnostic UMPL framework that leverages teacher-student uncertainty to refine pseudo labels and improve generalization with unlabeled data.

We explore the application of video generation models in controllable image generation tasks.

Introduce the first Open-World Hand-Object Interaction (HOI) Synthesis framework that can generate Long-horizon HOI sequences of Unseen Objects from Open-vocabulary instruction with 3D Multimodal Large Language Model.

We propose Equivalent Transformation Mechanism with KKT-Multiplier Regularization for solving SemiUOT and UOT

ALTo is a novel framework leveraging adaptive-length tokens for multimodal mask generation, based on MLLM

We propose a novel safety-driven unlearning framework for diffusion models that can better maintain unlearning performance after downstream fine-tuning.

Code Graph Models (CGMs) innovatively integrate both semantic and structural information from code repositories into LLMs, enabling effective repository-level coding tasks without relying on agents or closed-source models.

We propose InfiFPO, a novel model fusion method for preference alignment that integrates multi-source probability information to enhance LLM performance, outperforming existing approaches across 11 benchmarks.

We propose Fuz-RL, a fuzzy logic-based framework that enhances robustness of safe reinforcement learning by measuring and balancing uncertainties in a non-additive way, achieving better safety and performance under uncertainty.

Learning to forecast in the transformed domain improves forecasting performance.

This paper proposes a Multi-Model Contrastive Decoding algorithm and uses a dynamic hallucination detection mechanism in the generation phase to significantly reduce the model's hallucinations.

Training LLMs to combine reasoning with external knowledge retrieval via RL without any supervised data on reasoning steps.

Learning the individualized treatment effects under the assumption of rank preservation.

We propose OmniZoom, a universal plug-and-play framework for cross-device smooth zoom interpolation.

Enabling Instructional Image Editing in Large Scale Diffusion Transformer

To tackle the complex yet highly practical challenges, we propose a Theory-Driven Label-Specific Representation (TDLSR) framework.

VideoUFO is the first dataset curated in alignment with real-world users’ focused topics for text-to-video generation.

DataSIR, a multi-format sensitive information benchmark dataset, including 1,647,501 samples, designed to evaluate the performance of different models in recognizing sensitive information under evolving data leakage techniques.

ZPressor is an architecture-agnostic module that compresses multi-view inputs for scalable feed-forward 3DGS.

This paper introduces a novel module, IGAD, which can be integrated with reconstruction-based methods to address over generation and performance balance issues from a manifold perspective.

Our work effectively bridges Hebbian principles with explicit representation learning objectives, demonstrating considerable potential and biological plausibility.

We design an efficient data selection approach to construct a robust training mixture for visual instruction tuning in large multimodal models.

we introduce ThinkSound, a framework that utilizes Chain-of-Thought reasoning to systematically break down audio generation for videos into a step-by-step interactive process.

HFTP aligns syntactic processing representations in both large language models (LLMs) and the human brain

We present PLANA3R, a zero-shot method for metric Planar 3D Reconstruction from unposed two-view images.

We propose RepBlend, a multimodal dataset distillation framework that mitigates modality collapse via representation blending and symmetric projection matching, achieving superior performance and generalization.

We introduce an MLLM-based weather forecast quality analysis method, RadarQA, integrating key physical attributes with detailed assessment reports.

A novel Fast-Slow Thinking framework that dynamically adapts reasoning depth based on question characteristics.

SIU3R: the first alignment-free framework for generalizable simultaneous understanding and 3D reconstruction from unposed images.

To enhance the reliability of the reward model for current policy improvement, we have developed the Proximal Policy Exploration (PPE) algorithm to increase the coverage of the preference buffer in areas close to the near-policy distribution.

This paper presents MRSAudio, a large-scale multimodal recorded spatial audio dataset with refined annotations, designed for spatial audio generation and understanding tasks, along with its benchmarks.

we propose a framework propagating contextual information with geometric groups and semantic groups.

A novel semi-supervised learning paradigm that unifies view-wise co-training, meta-learned supervision, and adversarial perturbation through a structured triadic game.

Video Concept Transfer via Progressive Spatio-Temporal Decomposition

A progressive instruction evolution framework that iteratively unlock LVLMs' reasoning ability

We present Pixel-Perfect Depth, a monocular depth estimation model based on pixel-space diffusion transformers.

Based on a newly discovered "free lunch" in voxel labels, VoxDet reformulates 3D semantic scene completion as dense object detection using a VoxNT trick.

A pioneering multimodal LLM framework for table modality.

We propose a framework called Guard to enhance the robustness of a broad class of 1-consistent learning-augmented caching algorithms, while preserving 1-consistency and incurring low additional computational overhead.

A unified framework extending virtual try-on to any type of wearable object in mask-free manner.

We propose Refer-Judge, a Jury-and-Judge Chain-of-Thought framework that leverages MLLMs to uncover toxic data in 3D visual grounding

Make existing 3D generative models orientation aligned for orientation estimation and efficient insertion.

This study presents the first comprehensive investigation into model merging and data mixture strategies for constructing large language models (LLMs) aligned with the 3H principles (Harmlessness, Helpfulness, Honesty).

This paper reveals that backdoored LLM are more susceptible to new trigger injection during in-context learning (ICL) and proposes a targeted ICL-based lightweight backdoor detection framework for black-box LLMs used for generative tasks.

We propose a novel perspective to understand AI-generated image detection through the lens of OOD detection and introduce a generalizable training-free generated image detection method.

We propose FairDD to mitigate the bias in condensed datasets with a promising tradeoff between accuracy and fairness.

We improve the scalability of pathology foundation models through multi-scale vector quantization (MSVQ) to keep more patch tokens with compression, achieving performance comparable to other SOTA methods with significantly less WSI data.

we established optimal risk bound of $1/(\gamma^2 n)$ for GD with deep ReLU networks

A controllable federated group-fairness calibration framework that achieves global and local fairness in multi-class classification with theoretical guarantees.

HoliTom introduces a training-free holistic outer-inner token merging framework for video LLMs, significantly accelerating inference with negligible performance degradation.

We introduce a visual knowledge dataset aiming to evaluate current models' latest visual knowledge seeking and updating capabilities.

We introduce MS-Bench, a benchmark for ancient manuscript analysis, revealing how LMMs handle philological tasks, highlighting the potential for future human-AI collaboration.

We propose a confidence-separation self-distillation strategy and train SNNs using rate coding, achieving excellent performance with minimal training overhead.

Loong-X enables hands-free image editing using multimodal neural signals, achieving performance comparable to text-driven methods by combining BCIs with the proposed diffusion-based generative methods.

MoPFormer: Motion-Primitive Transformer for Wearable-Sensor Activity Recognition

FPSAttention is a training-aware FP8 quantization and sparsity co-design for video diffusion models that achieves up to 4.96× speedup without quality loss by aligning 3D tile granularity, denoising-step adaptation, and hardware-efficient kernels.

OLinear is a linear-based time series forecasting model that outperforms state-of-the-art Transformer-based methods with significantly higher efficiency.

Given unposed sparse-view images and an arbitrary style image, our method predicts stylized 3D Gaussians in less than a second using a feed-forward network.

We introduce the Remote Sensing Change Caption (RSCC) dataset, a large-scale benchmark comprising 62,315 pre-/post-disaster image pairs with detailed change captions.