This study compares Function Calling and Model Context Protocol architectures in LLM agents, revealing architecture significantly influences attack success rates, with chained attacks reaching 91-96%.

Tarek Gasmi, Ramzi Guesmi, Ines Belhadj et al.

Tora2 employs decoupled personalization and gated self-attention to enable multi-entity video generation with detailed appearance and motion control, outperforming previous methods.

Zhenghao Zhang, Junchao Liao, Xiangyu Meng et al.

GTA1 employs test-time scaling and RL-based coordinate prediction, achieving state-of-the-art GUI interaction accuracy with 83.1% on ScreenSpot-Pro.

Yan Yang, Dongxu Li, Yutong Dai et al.

PaddleOCR 3.0 introduces PP-OCRv5, PP-StructureV3, and PP-ChatOCRv4, enhancing multilingual recognition and document understanding.

Cheng Cui, Ting Sun, Manhui Lin et al.

OpenWorldSAM extends SAM2 with multi-modal embeddings for open-vocabulary segmentation, training only 4.5M parameters, achieving SOTA results.

Shiting Xiao, Rishabh Kabra, Yuhang Li et al.

Extending Diffusion Policy for multi-task robot models (LBMs), showing pretraining boosts success and robustness.

TRI LBM Team, Jose Barreiros, Andrew Beaulieu et al.

Gemini 2.5 Pro achieves SOTA in long multimodal reasoning with 3-hour video support using advanced Transformer architecture.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al.

ABench-Physics evaluates LLMs' physical reasoning via high-difficulty dynamic problems, revealing significant performance gaps.

Yiming Zhang, Yingfan Ma, Yanmei Gu et al.

DreamVLA predicts dynamic regions, depth, and semantics to build a perception-prediction-action loop, boosting robot manipulation success to 76.7%.

Wenyao Zhang, Hongsi Liu, Zekun Qi et al.

Introduced J1-ENVS and J1-EVAL for dynamic legal environment assessment; models scored below 60%, highlighting procedural gaps.

Zheng Jia, Shengbin Yue, Wei Chen et al.

MemOS introduces a system-level memory management framework for LLMs, significantly improving long-term memory, personalization, and knowledge consistency.

Zhiyu Li, Chenyang Xi, Chunyu Li et al.

EvoAgentX integrates TextGrad, AFlow, and MIPRO for automated multi-agent workflow evolution, achieving up to 20% performance gains.

Yingxu Wang, Siwei Liu, Jinyuan Fang et al.

Conformal Information Pursuit (C-IP) leverages prediction set sizes to estimate uncertainty, optimizing multi-round queries for large language models, improving efficiency and interpretability.

Kwan Ho Ryan Chan, Yuyan Ge, Edgar Dobriban et al.

Point3R uses explicit spatial pointer memory for streaming dense 3D reconstruction, achieving SOTA performance with low training costs.

Yuqi Wu, Wenzhao Zheng, Jie Zhou et al.

LiteReality combines scene understanding and model retrieval to convert RGB-D scans into compact, realistic, and interactive 3D scenes, suitable for AR/VR and digital twins.

Zhening Huang, Xiaoyang Wu, Fangcheng Zhong et al.

DeSTA2.5-Audio employs self-generated cross-modal alignment, using 5 million audio-text pairs from 50 datasets, overcoming catastrophic forgetting and achieving state-of-the-art results.

Ke-Han Lu, Zhehuai Chen, Szu-Wei Fu et al.

DexVLG predicts dexterous grasp poses using single-view RGBD input, achieving over 76% success rate.

Jiawei He, Danshi Li, Xinqiang Yu et al.

Proposes ACE estimator leveraging higher-order cumulants for robust causal effect estimation under non-Gaussian noise, outperforming DML.

Jikai Jin, Lester Mackey, Vasilis Syrgkanis

Kwai Keye-VL is an 8B multimodal model optimized for short-video understanding, leveraging large-scale video datasets and innovative training strategies.

Kwai Keye Team, Biao Yang, Bin Wen et al.

Selecting diverse, challenging reasoning traces from a strong teacher model enhances student model reasoning, outperforming existing datasets by 3-5% on benchmarks.

Yang Li, Youssef Emad, Karthik Padthe et al.