We propose a test-time adaptation method to identify and remove non-causal features.

We develop efficient algorithms for non-uniformly sampling over directed acyclic graph structures, and use these along with results from online learning, to develop efficient algorithms for agnostically-learning Bayes nets in KL divergence.

We present Auto-Connect, a method that establishe a new standard for automatic character rigging and skinning, offering a more reliable and effective solution for animation pipelines

We propose UtilGen, a novel utility-centric data augmentation framework which adaptively optimizes data generation based on downstream task feedback to produce high-utility training data.

A multi-level patch-instance-zone 3D representation model with a hierarchical dynamic online update strategy for embodied navigation.

an importance-sampling-based method to mitigate over-optimization in Direct Alignment Algorithms for language model alignment

We analyze the flow of tokens across attention layers and use these insights to enhance performance of Transformers.

We present Omni-Mol, a multitask molecular model that can solve any-to-any modality tasks

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

We investigate an efficient and effective context-window scheduling method for language model pretraining.

We introduce ObCLIP, a cloud-device hybrid oblivious image generation framework that achieves private Stable Diffusion inference (protecting the sensitive attributes in user prompts) with an enhanced efficiency-utility trade-off.

We propose a tree-sliced framework to the Partial Transport setting by observing that the Partial Transport problem on tree-metric spaces can be reformulated as a standard Optimal Transport problem.

Our paper introduces Semantic Surgery, a novel training-free method that dynamically manipulates text embeddings before diffusion to achieve concept erasure in text-to-image models while preserving image quality and locality.

We introduce MMOral, the first large-scale multimodal instruction dataset and benchmark tailored for panoramic X-ray interpretation. MMOral-Bench is a comprehensive evaluation suite covering five key diagnostic dimensions in dentistry.

We propose Distil-E2D for event-based monocular depth estimation, using dense synthetic depths from foundational models, a confidence-guided calibrated loss for label alignment, and a novel architecture to improve encoder representation.

We propose SAS to simulate larger attention head numbe and hidden size per head for better performance, keeping the original model size.

We introduce a novel embodied VLM agent with a VLM fine-tuned by agentic data synthesis for open-world mobile manipulation, unifying scene understanding, state tracking, and action generation for state-of-the-art results.

BC-LLM is a novel procedure that integrates Large Language Models into a Bayesian framework for concept discovery to achieve better predictive performance, converge faster to relevant concepts, and provide rigorous uncertainty quantification.

Talk2Event is a new benchmark for attribute-aware visual grounding from event cameras.

Large Reasoning Model learns when to think via Decoupled GRPO.

We introduce VeriThinker, a simple yet effective approach for CoT compression.

Analogy-Based Multi-Turn Jailbreak Method for LLMs

We view offline optimization from the new lens of a distributional translation task which can be modeled with a generalized Brownian Bridge Diffusion process mapping between the low-value and high-value input distributions.

We reveal that reasoning-oriented RL fine-tuning significantly increases the prevalence of hallucinations and propose a FSPO approach via reinforcement learning.

We propose HASTE, which combines holistic alignment (feature and attention) with early termination to accelerate diffusion transformer training by 28× while maintaining quality.

Motion-aware 4D Gaussian splatting enables pose-free reconstruction of dynamic scenes from monocular video with 5× faster training and superior rendering quality.

We propose a new paradigm to explore the best trade-off through mode connectivity for multi-task learning.

HyperGraphRAG introduces hypergraph-based knowledge representation to capture real-world n-ary relations, boosting answer accuracy, retrieval efficiency, and generation quality.

Generalized Video Anomaly Detection via Agentic AI Engineer

This paper measures and mitigates shortcut leaning in the misinformation detection.

A SoTA sequence parallelism for linear attention with a brand new collective communication.

This paper introduces Angular Steering, a robust and generalized method for fine-grained behavior control in language models, unifying and extending existing steering techniques through rotation in a feature-isolating subspace.

We develop a scalable second-order algorithm for pseudolikelihood-based partial correlation network estimation.

This work uses fractal features to reduce visual clues that are critical for privacy leakage while preserving high face recognition accuracy.

We propose EnzyControl, a method that integrates a lightweight, modular component into a pretrained motif-scaffolding model to enable functional and substrate-specific control over enzyme backbone generation.

Brain Harmony (BrainHarmonix) unifies T1 MRI + fMRI into 1D tokens and tops six brain-disorder benchmarks.

We propose a reward-centric approach to fast image generation that converts pretrained diffusion models to reward-enhanced few-step generators, without relying on complicated diffusion distillation losses or training images..

In this paper, we provide the first demonstration of a direct causal relationship between fairness and generalization. Furthermore, we show that both can be simultaneously improved by controlling for confounding factors.

We introduce EnerVerse, a generative robotics foundation model that constructs and interprets embodied spaces.

A novel time-aware data sharing framework that fairly decides and distributes model rewards to participating parties based on their value of data and time of participation to satisfy incentives.

To train an adaptive video tokenizer, we introduce probabilistic taildrop to inject visual complexity prior to the tokenizer and incorporate GRPO for post-training, which further boosts efficiency in a task-aware adaptive manner.

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

A reinforcement-learning post-training framework teaches LLM assistants to reason about contextual integrity, slashing inappropriate information disclosure while helping users complete their tasks.

A runtime verification framework of LLMs with domain knowledge.

We propose a novel influence-based adaptive sample borrowing approach to enhance treatment effect estimation in RCTs by effectively incorporating external controls.

We present the first theoretical convergence rates for constrained expected improvement, one of the most popular constrained Bayesian optimization methods.

How to employ the coherence between layers to design a PEFT method.

We study the approximation and generalization abilities of score-based neural network generative models

We propose the first defense framework designed for Audio-Language Models to defend against jailbreak attacks.

We propose a novel framework called GraphRTA, which conducts reprogramming on both the graph and model sides for open-set graph domain adaptation.

We analyze optimal budget allocation strategies for iterative bootstrapping with synthetic data to maximize model performance.

Towards Scene-Oriented Story Generation with Scene Planning and Scene Consistency

A Planning Representation and Paradigm Investigation of Vision-Language-Action Models

To incorporate reward shaping approach into multi-task reinforcement learning, we propose a Centralized Reward Agent based MTRL framework (CenRA) to share and transfer knowledge across multiple tasks.

Motion4D estimates consistent motion and semantics by integrating 2D models with explicit 4D reconstruction, outperforming SAM2 and state-of-the-art 2D tracking networks.

The paper proposes an LLM-guided knowledge graph traversal method that enhances retrieval efficiency and accuracy for complex queries while reducing costs.

Develop a test to check if model is obtained by fine tuning of another model

We propose a convolution-like linearization strategy that accelerates pre-trained diffusion transformers for ultra-resolution image generation.

NoisyRollout boosts VLM reasoning by mixing clean and noisy inputs during RL, improving generalization with no extra cost.

We propose FedWMSAM, a novel method that combines personalized momentum and dynamic SAM to achieve faster convergence and flatter minima in federated learning, effectively addressing client drift and data heterogeneity.

We develop "DroneAudioset", a dataset that provides audio recordings from drone-mounted microphones to inform noise suppression and sound classification techniques for drone-audition based human presence detection in search and rescue missions.

We study online learning of quantum states and prove that learning a pure state is as hard as learning a mixed state, id est that the rank of the matrix has no impact on the complexity of the learning problem.

A VLM method for vehicles and robots.

Teach multimodal models to think or not to think- this is a question.

We propose Approximate Domain Unlearning (ADU), a new task for selectively removing domain-specific knowledge from pre-trained vision-language models, and introduce two key techniques, domain disentangling and instance-wise prompting, to solve it.

We makes an early attempt at fully unsupervised LLM reasoning incentivization.

Generate high-performing LoRA parameters from prompts that is unseen in training.

Spiral is a new type of LiDAR generation model that enables semantic awareness and progressive diffusion.

We investigate Linear Mode Connectivity (LMC) in Mixture-of-Experts (MoE) architectures by analyzing their underlying permutation symmetries and proposing expert-matching algorithms that align independently trained MoEs to reveal LMC.

Concentric Ring Parallelism that reduces P2P communication volume and avoid inter-node communication.

We reduce training variance in equivariant generative models using a low-variance gradient estimator, improving stability and performance across molecular, crystal, and protein generation tasks.

FedMGP introduces a multi-group text-visual prompt paradigm for federated learning that effectively balances personalization and generalization , achieving state-of-the-art performance with minimal communication parameters.

We propose a new plug-and-play optimization method from the perspective of symmetry teleportation for mainstream multi-task learning approaches.

A large scale Dataset of Intracranial Aneurysm geometries and the associated hemodynamics simulation data.

ProteinConformers is a novel large benchmark dataset enabling better evaluation of protein structure generators by capturing realistic conformational diversity and plausibility and proposing new assessment metrics.