HoloV prunes 88.9% visual tokens in MLLMs while retaining 95.8% performance via spatial visual holistic context retention.

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 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.

We propose a novel AD pretraining framework specially designed for learning robust and discriminative pretrained representations for industrial anomaly detection.

The first animated QR code generation method.

We present PhysDiff-VTON via physics modeling and trajectory optimization for better virtual try-on performance.

We propose an adaptive, parameter-free accelerated method for local Hölder smooth convex optimization with optimal rates and no need for Lipschitz or target accuracy tuning.

We propose a new paradigm to accelerate different components of the VLA model and achieved state-of-the-art performance.

We ddvance vision-language action models via introducing comprehensive knowledge prediction.

In this paper, we decompose the reward value into prompt-free reward and prompt-related reward from a information-theoretic perspective, and use the former to guide reward training.

We propose ParaStep, a step-wise parallelization method based on a reuse-then-predict mechanism that parallelizes diffusion inference by exploiting adjacent-step similarity, reducing latency with minimal quality loss.

To address inefficiency from excessive visual tokens in LVLMs, we propose an information-flow perspective revealing dynamic redundancy emergence and introduce a method aligned with the model’s inherent behavior, outperforming all existing approaches.

We introduce an unsupervised method for post-training multi-modal large language models using implicit reward signals from majority voting based on GRPO.

PANTHER is a hybrid framework that combines self-supervised generative pretraining with lightweight discriminative modeling to enable real-time fraud detection in large-scale payment platforms, delivering a 38% improvement in online performance.

For deductive formal problem-solving (D-FPS), we propose a method and a data generation pipeline.

A report of the Big ANN competition in NeurIPS'23

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.

Training a new reasoning paradigm of LLMs explicitly contains meta-thinking in a multi-agent and multi-turn setting with RL

This paper presents a systematic pipeline for improving video generation with human feedback, including a large-scale preference dataset, a video reward model, and three alignment algorithms for flow matching models.

In this paper, we propose a novel Cascade Adaptive Self-Speculative Decoding (CAS-Spec) algorithm which constructs speculative draft models by leveraging dynamically switchable inference acceleration (DSIA) strategies

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

We propose a novel Dynamic Separator Configuration (DynSep) method that models separator configuration in different rounds as a reinforcement learning task, making decisions based on an incremental triplet graph updated by iteratively added cuts.

A reliable driving world model that can be accurately steered with various actions with realistic visual simulation.

We propose a structured denoising diffusion model, StruDiCO, which incrementally constructs solutions through step-wise variable selection.

This paper proves that specific frequency components of convolutional filters are robust to fine-tuning and uses them as neural network fingerprinting.

An efficient one-step diffusion model for real-world vsr.

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

We propose a new method that reduce the sampling steps of image auto-regressive models to only 1

An end-to-end RL-based OARSMT construction method.

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

We introduce the first graph foundation model specifically designed for retrieval-augmented generation in large language models.

We develop a benchmark for the classic combinatorial optimization problems (TSP, ATSP, CVRP, MIS, MCl, MVC, MCut) with relevant datasets.

We propose KRIS-Bench, a benchmark designed to evaluate knowledge-based reasoning in instruction-based image editing models.

AGC-Drive is a large-scale real-world dataset designed to support collaborative perception between aerial UAVs and ground vehicles across diverse driving scenarios.

This paper introduces CorrectBench, the first comprehensive benchmark for systematically evaluating self-correction mechanisms in LLMs.

A fast version for feature conformal prediction.

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 a dual-view graph learning framework that detects label noise by capturing semantic discrepancies between node-level and structure-level predictions.

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

We introduce MM-Agent and MM-Bench to enable and evaluate LLM-powered end-to-end mathematical modeling, achieving superior performance on real-world problems and Finalist Award in MCM/ICM 2025.

A unified and scalable RL framework applicable to online, offline, and offline-to-online settings.

This paper proposes a feedback-based conformal prediction framework for trajectory optimization that dynamically adjusts prediction regions using realized trajectories, improving decision performance while preserving coverage guarantees.

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

We present RiboFlow, a flow-matching based model that co-designs RNA structures and sequences with high binding affinity to target molecules.

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.

A novel one-step diffusion model for human body restoration, efficiently handling human motion blur and generic noise in human images.

We propose Retrieval-Augmented Diagnosis (RAD), a holistic method that explicitly integrates disease-centered medical knowledge into multimodal diagnosis models.

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

We propose a novel framework for multi-style sketch generation

CoT-Diff tightly couples MLLM reasoning with diffusion to enable step-by-step 3D layout planning and dynamic spatial control within a single generation process.

We systematically examine the current state of RLVR and surprisingly find that it does not elicit fundamentally new reasoning patterns—revealing a gap between the potential of RL and the actual impact of current RLVR methods.

Large fixed context length limits exploration and introduces redundancy in MARL. We propose an adaptive context length optimization method with Fourier-based low-frequency truncation to improve long-term decision-making.

This paper introduces a correlated low-rank adaptation strategy within each paired adjacent convolution layer for PEFT ConvNets, which demonstrates better convergence and performance than the high-rank tuning approach on downstream vision tasks.

This study propose a comprehensive benchmark for evaluating ML-based SAT solvers on cryptographic problems.

We propose a 3D all-atom flow matching model with prior interaction guidance and a learnable atom number predictor for target-aware molecule generation.

We propose a prune-then-finetune paradigm to specialize time series foundation models in both levels of architecture and parameter weights, effectively improving downstream forecasting performance.

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

We present a unified neural solver for multi-task VRPs using mixture of specialized experts based on the state-decomposable MDP.

This paper presents a physics-based humanoid control framework, aiming to mastering highly-dynamic human behaviors such as Kungfu and dancing through physics-based motion selection and adaptive motion imitation.

We train a token-level neural router to let SLM following LLM reasoning paths by replacing only divergent tokens

We present H3-DDG, a model that predicts binding free energy changes by modeling higher-order many-body interactions, achieving state-of-the-art performance in protein mutational effect prediction.

Reinforcement Learning Successfully Attempt for End-to-End Autonomous Driving in CARLA v2

This study reveals and quantifies key factors in visual token pruning from a geometric covering perspective, and proposes Multi-Objective Balanced Covering, which significantly accelerates MLLM reasoning with negligible performance loss.

This paper presents InstructSAM, a training-free framework that counts, detects, and segments objects in remote sensing imagery based on tailored instructions.

Hypergraph Message Passing Inspired by Particle System Theory

We propose VIKI-Bench and VIKI-R to advance visual-driven cooperation in embodied multi-agent systems.

More accurate RNA inverse design model

KV cache retrieval for large language models using nonlinear hashing function.

Segmentation-oriented Industrial Anomalies Synthesis

We present UniVF, a unified video fusion framework with multi-frame learning; VF-Bench, the first benchmark covering four video fusion tasks; and a unified spatial–temporal evaluation protocol with a new temporal consistency loss.

We propose a progressive consistency distillation framework that enhances the efficiency of MLLMs by significantly reducing computational cost while preserving strong performance.

We present MICS, a multi-model collaborative search strategy that facilitates the generation of effective step-by-step CoT data.

We introduce a RL framework to train LLM's reasoning and self-verification ability simultaneously.

We show that high-dimensional neural activity can arise from low-dimensional latent dynamics, both in RNNs and in the brain.

In this paper, we propose a fine-grained influence function framework to trace how training data on SFT phase shapes LLM reasoning in math and code tasks.

A lightweight, plug-and-play mapper to boost the performance of OVSS with minimal computational overhead

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

We propose a data selection method that leverages sparse, monosemantic neuronal activations learned via a sparse autoencoder to improve task-specific instruction tuning for large language models.

Streaming dynamic scene reconstruction with a novel 3d motion representation, accurate, fast, compact.

REArtGS

A graph-structured perceptron model with self-assembly capabilities.

We introduce PhyBlock, a progressive benchmark evaluating large vision-language models on physical understanding and spatial planning via robotic 3D block assembly tasks.

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

An Effective Levelling Paradigm for Unlabeled Scenarios

We introduce the concept of semantic orientation, representing the object orientation condition on open vocabulary language.

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

AdaTS introduces a dynamic soft contrastive learning framework for adaptive time series representations, leveraging frequency-domain similarities with dynamic temporal and ordinal instance contrasts.

An energy-based diffusion model regularized with the Fokker-Planck equation to improve the consistency between iid sampling and simulation

Tri-MARF, a novel tri-modal multi-agent framework, integrates 2D images, text, and 3D point clouds with specialized agents to enhance 3D object annotation, achieving superior accuracy, retrieval, and throughput.

LeVo generates high-quality songs that closely follow instruction by pairing paralled mixed and dual-track token prediction with three-stage training and DPO-based multi-preference alignment.

We introduce rStar-Coder to train advanced code reasoning LLMs, with our 14B model achieving comparable performance to QWQ-32B.

In this work, we repurpose AlphaFold3-like protein folding models for antibody sequence and structure co-design.

We propose a flexible realignment framework that enables efficient and controllable realignment of LLMs during both training and inference, addressing challenges in reasoning efficiency and personalized response balancing.

We propose Open-IRT, a vision-language-based test-time adaptation method for single-sample open-set streams, achieving state-of-the-art performance via polarity-aware OOD filtering and intermediate domain learning.

We propose a universal video grounding model based on MLLMs, which achieves superior accuracy, generalizability, and robustness.

An efficient PbRL method that mitigates overfitting and overestimation via dual regularization, enhancing feedback efficiency in both online and offline settings

We propose GenPO, which effectively incorporates invertible diffusion model into on-policy RL, and deals with the challenge of log-likehood computation in diffusion policies.

RetrievalAttention improves decoding speed and reduces GPU memory usage in Transformer-based LLMs by using pre-built, attention-aware KV vector indexes stored in CPU memory, achieving significant efficiency gains without compromising accuracy.

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

We propoed the first application of RLVR to directly enhance LLMs' proficiency in optimization modeling: SI-RL, achieves state-of-the-art on diverse public benchmarks

Hierachical Balance Packing, Efficient SFT

We establish a KKT Point Embedding Principle where max-margin solutions (KKT points) in homogeneous networks cleanly embed into wider networks formed by neuron splitting, a property preserved throughout gradient flow dynamics.

An LLM-based retrosynthesis planning agent trained end-to-end with Agentic Reinforcement Learning.

We introduce MMAR, a new benchmark designed to evaluate the deep reasoning capabilities of Audio-Language Models (ALMs) across massive multi-disciplinary tasks.