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.

This paper proves the convergence of learn to optimize.

We study the expression and limitations of transformers in performing in-context learning of functions governed by Markovian dynamics.

This paper explores the representation of robot actions in the frequency domain and introduces FreqPolicy, a novel frequency-domain autoregressive framework for hierarchical robotic action generation.

We present ShotBench, a new benchmark for evaluating VLMs' cinematography understanding, along with the ShotQA dataset and our ShotVL model, which achieves state-of-the-art performance over both strong open-source and proprietary baselines.

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 introduce PHYSICS, a dataset containing 16,568 high-quality physics problems spanning subjects and difficulty levels, and an improved evaluation framework combining rule- and model-based judgments to advance LLMs' physical reasoning.

BREEZE is a new zero-shot reinforcement learning framework that improves the stability, expressivity, and performance of pre-trained generalist policies, achieving state-of-the-art results on benchmark tasks.

We compare RL methods DPO & GRPO for image generation, showing their strengths & how rewards affect generalization. Explores scaling for better CoT-based synthesis.

HiFlow is a training-free and model-agnostic framework to unlock the resolution potential of pre-trained flow models.

Imagine360 creates high-quality, immersive 360 videos from perspective video anchors.

A Comprehensive Benchmark of MLLM for Endoscopy Analysis

This paper introduces a simple and scalable semi-off-policy reinforcement learning method, i.e., SOPHIA, to enhance LVLMs’ ability to perform visual slow-thinking reasoning.

We propose an information-preserving and data-free pipeline for ultra-efficient delta compression, achieving ultra-high compression ratios and strong performance across various models.

A data construction pipeline and a DiT framework for subject-driven video customization under multimodal control conditions

A memory framework for video world models that improves long-term consistency

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.

We learn offline meta-policies from natural language supervision with contrastive language-decision pre-training, aligning text embeddings to comprehend environment dynamics.

We propose FairImagen that debiases text-to-image models by post-processing prompt embeddings, improving fairness across gender and race without retraining the model.

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

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

We propose GRiD, a framework that grounds reasoning in knowledge-enhanced dependency graph, ensures logical consistency, and improves reasoning accuracy across various benchmarks.

We meta-learn a transformer based in-context learning fMRI visual cortex encoder which can adapt to new human subjects without any fine-tuning

We present COS3D, a new collaborative prompt-segmentation framework that effectively contributes to integrating complementary language and segmentation cues throughout the entire pipeline.

We propose Fira, a plug-and-play memory-efficient training framework of LLMs to enable full-rank training consistently under the low-rank constraint.

We understand oversmoothing through the lens of signed graphs and propose a plug-and-play method to address it.

MedChain, a dataset of 12,163 clinical cases that mimics real-world medical practice through personalization, interactivity, and sequentiality. MedChain-Agent, an AI system with case-based learning and feedback mechanisms.

This paper proposes conflict-aware multi-objective training strategies for multilingual automatic speech recognition and speech translation by selectively aligning gradients from conflicting layers to enhance efficiency and performance.

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

We systematically introduce AND, OR, and ADDER gates and analyze the faithfulness and completeness via these gates, and we also propose a framework to fully discovery these gates and obtain some findings.

We propose NS-EGNN, which is capable of modeling non-stationary physical dynamics with equivariance.

We propose TreeSynth, a tree-guided subspace-based data synthesis approach, achieving superior data diversity, model performance, robust scalability, and data balance efficacy.

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

We propose Flow-GRPO, the first method to integrate online RL into flow matching models, significantly enhancing text-to-image generation performance.

A novel method to efficiently fine-tune Large Language Models in hyperbolic space, unlocking their latent tree-like structures and significantly boosting complex reasoning performance by adapting directly on the hyperbolic manifold.

We propose a value gradient matching formulation for reward finetuning/alignment for flow matching models with the theory of optimal control, and empirically verify our method on the popular text-to-image flow matching model StableDiffusion3

Hierachical Balance Packing, Efficient SFT

We propose Adaptive Classifier-Free Guidance (A-CFG), which dynamically re-masks low-confidence tokens for a more targeted unconditional input in iterative models.

Instruction-based Video Editing Dataset

LithoSim introduces lithography simulation benchmark with >4 million rigorously curated input-output pairs, integrating optical variations, mask corrections, and process dynamics to establish a unified evaluation flow for ML-based simulation.