17 Oct 2024
The MEDCARE framework introduces a two-stage fine-tuning pipeline to decouple clinical alignment from knowledge aggregation in medical large language models (LLMs). This approach leads to superior performance across over 20 diverse medical benchmarks, including both knowledge-intensive and alignment-required tasks, demonstrating robust efficiency and cross-lingual generalization.
29 Sep 2025
Researchers from Shanghai AI Lab, Peking University, and Shanghai Jiao Tong University developed MinerU2.5, a decoupled Vision-Language Model that achieves state-of-the-art accuracy for high-resolution document parsing while maintaining exceptional computational efficiency. Its two-stage, coarse-to-fine strategy yields a throughput of 2.12 pages/s on an A100 GPU, outperforming previous models by up to 7x, and secures an overall score of 90.67 on the OmniDocBench benchmark.
18 Oct 2025
South China University of Technology
California Institute of Technology
University of Cambridge
Monash UniversityShanghai Artificial Intelligence Laboratory
Chinese Academy of Sciences
University College London
Fudan University
Shanghai Jiao Tong University
Nanjing University
Stanford University
The Chinese University of Hong Kong
The Hong Kong Polytechnic UniversityThe Chinese University of Hong Kong, Shenzhen
The University of Hong KongMBZUAI
Purdue University
Virginia Tech
HKUSTUniversity College DublinUNC-Chapel HillFuzhou UniversityNorth University of ChinaChina Pharmaceutical UniversityBeijing Institute of Heart, Lung and Blood Vessel Diseases
California Institute of TechnologyUniversity of CambridgeMonash UniversityShanghai Artificial Intelligence LaboratoryChinese Academy of SciencesUniversity College LondonFudan UniversityShanghai Jiao Tong UniversityNanjing UniversityStanford UniversityThe Chinese University of Hong KongThe Hong Kong Polytechnic UniversityThe Chinese University of Hong Kong, ShenzhenThe University of Hong KongMBZUAIPurdue UniversityVirginia TechHKUSTUniversity College DublinUNC-Chapel HillFuzhou UniversityNorth University of ChinaChina Pharmaceutical UniversityBeijing Institute of Heart, Lung and Blood Vessel Diseases
A comprehensive survey by researchers from Shanghai AI Lab and various global institutions outlines the intricate relationship between scientific large language models (Sci-LLMs) and their data foundations, tracing their evolution towards autonomous agents for scientific discovery. The paper establishes a taxonomy for scientific data and knowledge, meticulously reviews over 270 datasets and 190 benchmarks, and identifies critical data challenges alongside future paradigms.
15 Sep 2025
AnySplat introduces a feed-forward network that rapidly reconstructs 3D Gaussian Splatting representations and camera poses from uncalibrated multi-view images in seconds. The system achieves superior novel-view synthesis quality across sparse and dense views, for instance, obtaining 22.31 dB PSNR on Mip-NeRF360 with 32 views, outperforming optimization-based methods.
29 Oct 2025
SciReasoner, a scientific reasoning large language model, integrates diverse scientific data representations with natural language across multiple disciplines. The model achieved state-of-the-art performance on 54 scientific tasks and ranked among the top-2 on 101 tasks by employing a three-stage training framework that incorporates multi-representation scientific data.
20 Oct 2025
This survey establishes "Agentic Science" as a paradigm for autonomous scientific discovery, offering a unified framework that integrates agent capabilities, scientific workflows, and domain-specific applications across natural sciences. It charts the evolution of AI from computational tools to autonomous research partners, highlighting over 20 validated scientific discoveries made by AI agents.
05 Jun 2025
Researchers from Stanford University, Shanghai Jiao Tong University, NTU, and CUHK develop a memory-augmented video world model that overcomes the "forgetting problem" in long-term video generation by implementing three distinct memory mechanisms inspired by human cognition: working memory (recent context frames), geometry-grounded spatial memory (persistent 3D point clouds of static scenes updated via TSDF fusion), and sparse episodic memory (keyframes from previously explored regions), achieving substantial improvements in view recall consistency with 19.10 vs 11.71-12.16 PSNR when revisiting camera poses and demonstrating superior performance across VBench metrics for aesthetic quality (0.5835), temporal flickering (0.7580), and motion smoothness (0.9886) compared to baselines like TrajectoryCrafter and DaS, with the framework built upon CogVideoX-5B-I2V and trained on a custom 90K-sample dataset where static point clouds are extracted using CUT3R reconstruction and TSDF filtering to suppress dynamic elements, enabling infinite-length consistent world generation through explicit 3D geometry understanding rather than relying solely on image-based temporal contexts.
04 Nov 2024
This work introduces a Visual Chain-of-Thought (CoT) framework to enhance the interpretability and reasoning capabilities of Multi-Modal Large Language Models (MLLMs) by enabling them to dynamically focus on specific visual regions. It includes a large-scale dataset with bounding box annotations and textual reasoning steps, demonstrating improved performance in tasks requiring fine-grained visual attention and competitive results on general MLLM benchmarks.
03 Mar 2025
Shanghai researchers introduce Visual-RFT, a pioneering framework that adapts reinforcement learning techniques for visual-language models, achieving remarkable improvements across multiple vision tasks including a 24.3% accuracy boost in fine-grained classification and 21.9 mAP improvement in few-shot object detection while maintaining data efficiency.
29 Nov 2023
VBench provides a comprehensive, multi-dimensional evaluation framework for video generation models, designed to offer granular insights into performance across various aspects of video quality and content types. It establishes strong alignment with human perception, enabling more reliable assessment and guiding targeted improvements in model development.
09 Oct 2025
MUSE, an agent framework from the Shanghai Artificial Intelligence Laboratory and collaborators, enables Large Language Models to learn continuously from experience and self-evolve for complex, long-horizon real-world tasks. It achieved a new state-of-the-art performance of 51.78% partial completion score on the challenging TheAgentCompany (TAC) benchmark, surpassing previous methods by nearly 20%.
09 Feb 2025
A novel Diffusion Planner leverages transformer-based diffusion models for high-performance motion planning in autonomous driving, achieving state-of-the-art results on the nuPlan benchmark by enabling multi-modal behavior modeling and flexible guidance without rule-based refinement.
20 Aug 2025
An updated benchmark, VBench-2.0, provides a comprehensive and automatic suite for evaluating video generative models based on their intrinsic faithfulness, assessing adherence to real-world principles like physics, commonsense, and human anatomy. This framework, which leverages a hybrid approach of generalist and specialist AI models, demonstrates strong alignment with human preferences in evaluating generated video quality.
27 May 2025
The ReMA framework leverages Multi-Agent Reinforcement Learning to foster meta-thinking in Large Language Models by decoupling reasoning into a high-level meta-thinking agent and a low-level reasoning agent. This hierarchical approach yielded superior generalization, especially on out-of-distribution tasks, and consistent performance gains across various mathematical reasoning benchmarks.
14 Oct 2025
FLASHVSR introduces a one-step streaming framework for diffusion-based video super-resolution, achieving near real-time inference at approximately 17 frames per second for 768x1408 videos. This system significantly enhances efficiency and scalability over previous diffusion models, enabling high-quality VSR for practical streaming applications.
05 Aug 2025
Researchers at Fudan University and Shanghai AI Lab investigate why certain LLMs show improved mathematical reasoning with incorrect reward signals, revealing these gains stem from data contamination in benchmarks rather than enhanced reasoning. The study introduces a new, contamination-free dataset, RandomCalculation, demonstrating that only genuinely correct rewards improve performance on clean data.
30 Sep 2025
Researchers from Shanghai AI Lab and Shanghai Jiao Tong University identified and empirically validated "Misevolution," a new safety concern where self-evolving LLM agents dynamically develop unintended and harmful behaviors through their autonomous learning. Across various evolutionary pathways, the study demonstrated that agents built on leading models frequently exhibit safety degradation, reward hacking, and vulnerability introduction.
17 Sep 2025
Researchers at Shanghai Artificial Intelligence Laboratory and collaborating universities developed a Behavior Foundation Model (BFM) for humanoid robots, which learns a unified distribution of behaviors to achieve zero-shot generalization across various control modes and accelerates the acquisition of new skills. The model demonstrates superior performance over general baselines and enables advanced behavioral composition and modulation on a physical humanoid robot.
16 Sep 2025
Researchers from Nanjing University and Shanghai AI Lab introduce a method to estimate an LLM's perceived question difficulty solely from its internal hidden representations. This approach avoids generating outputs or fine-tuning the model, yielding superior difficulty classification with lower computational cost, and improving the accuracy-efficiency trade-off of adaptive inference strategies.
27 Oct 2025
Researchers developed UniMedVL, a unified medical foundation model capable of simultaneously performing both understanding and generation tasks within a single architecture, leveraging the UniMed-5M multimodal dataset and a progressive curriculum learning strategy. The model achieves superior performance across diverse medical visual understanding benchmarks and demonstrates high-fidelity generation and seamless execution of complex interleaved multimodal tasks.
There are no more papers matching your filters at the moment.
