We propose a new error metric for constructing coresets in \$(k,z)\$-clustering with noisy data, leading to smaller coresets, stronger theoretical guarantees, and improved empirical performance compared to classical methods.

We propose SRPO, a reflection-aware RL method that significantly improves multimodal LLM reasoning by explicitly teaching self-reflection, outperforming state-of-the-art models on multiple benchmarks.

a novel approach to enable training high resolution sparse voxel diffusion model for 3D generation

We identify the output projection module as the key component enabling reasoning in LLMs, and propose Stethoscope for Networks (SfN) to diagnose and support this claim.

We introduce MotionRAG, a framework that improves video generation by retrieving motion from real videos and injecting it into diffusion models to create more realistic movement.

This paper proposes a fair federated learning method for handling client participation probability inconsistencies in heterogeneous environments

This paper enhances contrastive learning by introducing the variable similarity of augmented views.

We propose GOOD, a training-free framework that leverages off-the-shelf classifiers to guide diffusion models for generating diverse, informative OOD samples—improving outlier exposure without requiring external datasets or embedding alignment.

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.

This paper presents a simple OOD detection approach by mixing multi-layer representations of Transformer.

We propose a token- and task-wise MoE structure to in-context RL models, harnessing architectural advances of MoE to unleash RL's in-context learning capacity.

We propose Curriculum Abductive Learning (C-ABL), a method that explicitly leverages structural properties of domain knowledge bases to improve training efficiency, stability and reasoning accuracy in ABL.

A foundation model specifically designed for medical time series forecasting

We propose VoTA, a novel attack framework that exploits the tension between logical reasoning and safety objectives in VLMs by generating chains of images with risky visual thoughts, achieving significantly higher success rates than existing methods.

A challenging tri-modal reasoning benchmark for evaluating omni-language models.

We propose new optimistic online-to-batch conversions for acceleration and universality.

ScoreRS, a quality assessment model trained on carefully curated remote sensing vision-language preference data. ScoreRS effectively scores and filters vision-language datasets, improving model performance by selecting high-quality data.

This paper introduces CoPDT, a method of using one unified and adaptable DT model for multi-task (multi-budget or multi-constraint) offline safe RL.

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

We propose GUI-Actor, a VLM-based, coordinate-free GUI grounding method with an attention-based action head and verifier, achieving state-of-the-art results and strong generalization.

We propose a novel MOE-based architecture that combines the detector and the visual foundation model for cross-domain few-shot object detection.

A model-based reinforcement learning framework that effectively extracts task-relevant information from noisy observations by leveraging conditional dependence.

We present the first long-term rehearsal learning approach, which demonstrates favorable properties such as variance reduction and optimality.

We propose a novel system identification framework for sim-2-real transfer in reinforcement learning that combines Diffusion Evolution with Adversarial Learning (DEAL) to iteratively infer physical parameters with limited real-world data.

We propose LongBioBench for controllable evaluation on Long-Context Language Models

We investigate last-iterate convergence of Regret Matching$^+$ variants in games satisfying the weak Minty variation inequality.

We propose KORGym, a dynamic, game‐based benchmark offering over 50+ interactive tasks with RL support for multi‐turn LLM reasoning evaluation, and validate its effectiveness through extensive experiments, revealing several key insights.

We introduce EvaLearn, a pioneering benchmark designed to evaluate LLMs on their learning capability and efficiency in challenging tasks.

We introduce TwinMarket, an LLM-based multi-agent framework that simulates socio-economic systems by modeling how individual behaviors interact to produce emergent market dynamics like bubbles and crashes.

We propose a 3D VLA model that aligns the input and output within a shared 2D space in both pre-training and fine-tuning which enables high data efficiency and achieves impressive performance in both basic and generalization settings.

We introduce the first learning-augmented streaming algorithms for Correlation Clustering on both complete and general graphs, improving the best-known space-approximation tradeoffs.

A novel two‑stage Direct Preference Optimization framework that enables short‑context vision‑language models to robustly understand ultra‑long videos without any long‑video annotations.

A novel value decomposition framework of Continued Fraction Q-Learning (QCoFr) is proposed to model rich cooperation for multi-agent reinforcement learning without combinatorial explosion.

LoTA-QAF uses ternary adaptation for fine-tuning quantized LLMs, enabling the lossless merging of adaptation into quantized weights.

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 enhances the fully decentralized cooperative multi-agent reinforcement learning from a context modeling perspective.

A practical and super-fast QAT method for general vision applications; source for recognition and detection will be released.

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

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

We show how perceived post-selection bias distorts strategic effort in merit-based selection, leading to disparities. Our model quantifies interventions to reduce inequity by adjusting selectivity and perceived valuation gaps.

This paper introduces a nove metric (REG) for evaluating the reasoning efficiency of LRMs and a reinforcement learning method (REO-RL) that significantly reduces reasoning redundancy while maintaining accuracy.

We present the first parameter-free last-iterate convergence of Counterfactual Regret Minimization algorithms.

We introduce a novel object selection mechanism to allow sim-trained policies to rapidly adapt to real-world visual perturbations.

We systematically investigate the design space and scaling property of native Multimodal Large Language Models and introduce a novel MLLM that achieves competitive performance against existing MLLMs.

We design a training-free neural architecture search method for Mamba2.

This study introduces VADB, the largest video aesthetic database with 10,490 videos, and VADB-Net, a novel framework that outperforms existing models in video aesthetic assessment.

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

We propose a novel continual learning method for language models, built upon the low-rank adaptation framework.

We propose a dual-driven (data and knowledge) approach for symbolic PDE discovery that exhibits superior performance in terms of noise robustness and hyperparameter stability.

We present FreeControl, a training-free method for structural control in diffusion models that achieves strong semantic alignment with less than 2\% additional inference cost.

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.

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.

This paper presents SongBloom, the first unified autoregressive diffusion model for long-form song generation, achieving state-of-the-art performance compared to both commercial and non-commercial methods.

We introduce the first theoretical framework for analyzing LLM reasoning errors, and bridge two typical sampling-based test-time scaling methods to achieve both low error and fast convergence.

We propose SignViP which incorporate multiple fine-grained conditions with a discrete tokenization paradigm for improved generation fidelity.