This paper proposes a data-free dual orthogonal projection framework to perform continual model merging.

This paper proposes Panacea, a post-fine-tuning method that mitigates harmful fine-tuning in large language models, maintaining safety alignment without sacrificing performance across different tasks and models.

We propose a training-free projection-based continual merging method that efficiently combines models incrementally.

An effective parameter-efficient model merging method for multimodal large language models from the perspective of direction robustness in low-rank space

In this paper, we propose Value-Guided Decision Transformer (VDT), which employs progressively optimized value functions to guide the Decision Transformer (DT) in making optimal decisions.

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.

This article proposes the first refined dataset, A Fine-Grained Global Terraced Parcel and Boundary Dataset(GTPBD), covering typical terraced field distribution areas worldwide. It is innovative and supports multi task evaluation.

MM-OPERA is a benchmark of 11,497 open-ended association tasks (Remote-Item and In-Context) with explicit multi-hop reasoning and LLM-as-a-Judge process-reward evaluation to assess LVLMs' convergent and divergent associative thinking.

A robust watermarking scheme for intellectual property protection in diffusion models.

We present a prompt based multimodal semantic segmentation on the basis of pertained single-modality RGB model

A Unified Framework for Dynamic Gaussian Splatting from Defocused and Motion-blurred Monocular Videos

Our empirically and theoretically informed method, which treats diversity as a reward, achieves new SOTA average performance across 7 benchmarks on SOTA LLMs with domain-undetermined data.

A Comprehensive Benchmark of MLLM for Endoscopy Analysis

SeasonBench-EA, a multi-resolution, multi-source benchmark dataset for seasonal climate prediction over East Asia, supporting both machine learning-based forecasting and ensemble post-processing.

This paper introduces CERES that uses dual-modal causal intervention (backdoor adjustment for language bias and front-door adjustment with vision-depth fusion for visual bias) to achieve robust, SOTA egocentric referring video object segmentation.

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

We propose CSBrain, a Cross-scale Spatiotemporal Brain foundation model for generalized EEG decoding, which effectively captures diverse spatiotemporal neural patterns and achieves state-of-the-art performance on 11 tasks across 16 public datasets.

VLMs lack deep causal reasoning. CF-VLM, via counterfactuals & novel training, boosts causal logic. It surpasses SOTA in reasoning/generalization, cuts hallucinations, aids real-world VLM use.

We propose a principled framework Bayesian Data Scheduler (BDS), which is an adaptive tuning-stage defense strategy against harmful fine-tuning with no need for attack simulation.

Proposing a zero-shot cross-task manipulation generalization benchmark and a novel generalizable VLA method.

Efficient Long Context Fine-tuning through Dynamic Data Scheduling

Sequential updates with Parallel aggregation for Enhanced Federated Learning under Category and Domain Shifts

We propose GAM-Agent, a game-theoretic multi-agent framework where visual and logic agents debate via structured communication and uncertainty control, boosting VLM performance, robustness, and interpretability. It is modular, scalable, and general.

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.

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

We propose DiffLiG, a graph-based model that addresses station heterogeneity and oversmoothing via spatiotemporal encoding and diffusion-based ensemble refinement for accurate, structure-aware precipitation interpolation.

We have decoupled the math atomic capabilities of large language models and explored their interaction relationships in mathematical reasoning tasks.

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

We distill commonsense spatial priors from large language models into a flow-based generative model to help agents imagine unseen parts of indoor scenes for Object Goal Navigation task.

A novel framework for spike camera reconstruction

we introduce the use of the memory buffer into the video stereo matching task, aiming to construct a memory buffer that is both compact and informative.

We propose AlphaDecay, a per-module weight decay method guided by heavy-tailedness, improving large language model performance.

DynaPipe dynamically redistributes layers and uses asynchronous coordination to balance computation during LLM inference, significantly reducing latency and outperforming existing pipeline parallelism systems.

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.

A Reasoning Benchmark on Visual Perception

A novel approach; Very practical

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.

We present GLARE, a reinforced active learning framework that improves virtual screening efficiency by dynamically optimizing molecular selection while balancing chemical diversity, biological relevance, and computational constraints.

We introduce HBLLM, a wavelet-enhanced high-fidelity $1$-bit post-training quantization method for Large Language Models (LLMs).

We propose MindGYM, a thinking-centric data synthesis framework that injects cognitive traits into QA generation, enabling language and vision-language models to self-synthesize high-quality, low-variance data for efficient fine-tuning.