Top 15 AI Papers: November 2025 Arxiv Daily
Stay up-to-date with the latest advancements in Artificial Intelligence! This article summarizes the top 15 papers featured in the Arxiv Daily mailing on November 27, 2025, curated by yanghlll and ArxivDaily-Haolin. For a better reading experience and access to more papers, check out the Github page.
Unified AI Models: The Quest for General Intelligence
The pursuit of unified AI models is a significant trend in the field, aiming to create systems that can handle a wide range of tasks and modalities. This section highlights research focused on building more versatile and capable AI.
In the realm of unified models, a groundbreaking paper titled GVLM: Geometry Grounded Vision Language Model with Unified 3D Reconstruction and Spatial Reasoning, published on November 26, 2025, presents a novel approach to vision-language modeling. This model integrates 3D reconstruction and spatial reasoning, allowing for a more comprehensive understanding of visual scenes. The code for this project is available on Github, making it easier for researchers to build upon this work.
Another noteworthy contribution is BoundingDocs: a Unified Dataset for Document Question Answering with Spatial Annotations, also released on November 26, 2025. This dataset aims to facilitate research in document understanding by providing spatial annotations, which are crucial for tasks like information extraction and question answering. Datasets like BoundingDocs are essential for training and evaluating AI models that can process and understand complex documents.
UniChange: Unifying Change Detection with Multimodal Large Language Model, published on the same day, explores the use of multimodal large language models for change detection. This research is particularly relevant in applications like remote sensing and surveillance, where identifying changes in the environment is critical. By leveraging the power of large language models, UniChange offers a promising approach to this challenging problem.
XYZCylinder: Towards Compatible Feed-Forward 3D Gaussian Splatting for Driving Scenes via Unified Cylinder Lifting Method introduces a novel method for 3D Gaussian Splatting, specifically tailored for driving scenes. This technique, which falls under the category of Feed-Forward 3D Gaussian Splatting, uses a Unified Cylinder Lifting Method to achieve compatibility. The project page, available at https://yuyuyu223.github.io/XYZCYlinder-projectpage/, provides further details and resources for those interested in this approach.
The paper CtrlVDiff: Controllable Video Generation via Unified Multimodal Video Diffusion, spanning 27 pages with 18 figures and 9 tables, presents a method for controllable video generation. This approach leverages unified multimodal video diffusion, allowing for greater control over the generated content. The project page, accessible at https://tele-ai.github.io/CtrlVDiff/, offers additional insights and examples of this technology.
A Unifying View of Linear Function Approximation in Off-Policy RL Through Matrix Splitting and Preconditioning provides a theoretical framework for understanding linear function approximation in off-policy reinforcement learning. This work, accepted for a spotlight presentation (top 3% of papers) at NeurIPS 2025, offers valuable insights into the foundations of reinforcement learning algorithms. Theoretical advancements like this are crucial for developing more robust and efficient RL systems.
Saliency-R1: Incentivizing Unified Saliency Reasoning Capability in MLLM with Confidence-Guided Reinforcement Learning explores the use of reinforcement learning to enhance saliency reasoning in multimodal large language models (MLLMs). The main text of the paper, excluding references, is 8 pages long and includes 4 figures, while the supplementary materials consist of 9 pages and 10 figures. This research highlights the importance of saliency reasoning in MLLMs, which enables them to focus on the most relevant parts of an image or video.
UniGame: Turning a Unified Multimodal Model Into Its Own Adversary presents a novel approach to training unified multimodal models by making the model compete against itself. This adversarial training technique can lead to more robust and generalizable models. Self-adversarial training is a powerful technique for improving the performance of AI systems.
A Unified Understanding of Offline Data Selection and Online Self-refining Generation for Post-training LLMs focuses on post-training techniques for large language models, specifically offline data selection and online self-refining generation. These methods aim to improve the performance of LLMs without requiring extensive retraining. Post-training techniques are essential for adapting LLMs to specific tasks and domains.
MUSE: Manipulating Unified Framework for Synthesizing Emotions in Images via Test-Time Optimization introduces a framework for synthesizing emotions in images using test-time optimization. This approach allows for fine-grained control over the emotional content of generated images. The ability to manipulate emotions in images has applications in areas like art, advertising, and entertainment.
The paper CartoonSing: Unifying Human and Nonhuman Timbres in Singing Generation explores the challenging task of generating singing voices with both human and nonhuman timbres. This research pushes the boundaries of audio synthesis and has potential applications in music production and virtual entertainment. Generating realistic and expressive singing voices is a complex problem in AI.
CameraMaster: Unified Camera Semantic-Parameter Control for Photography Retouching presents a method for controlling camera parameters semantically, enabling more intuitive and powerful photography retouching. This technology has the potential to simplify and enhance the image editing process. User-friendly interfaces for image manipulation are a key area of development in AI.
Primal: A Unified Deterministic Framework for Quasi-Orthogonal Hashing and Manifold Learning introduces a unified framework for hashing and manifold learning, providing a theoretical foundation for these techniques. This research contributes to the understanding of dimensionality reduction and similarity search. Theoretical frameworks are essential for guiding the development of new algorithms and methods.
A Psychology-based Unified Dynamic Framework for Curriculum Learning, accepted for publication in Computational Linguistics, presents a curriculum learning framework inspired by psychology. The code for this project is available at https://github.com/nd-ball/cl-irt. Curriculum learning, which involves training models on progressively more difficult tasks, is a promising approach to improving the learning efficiency of AI systems.
Finally, iMontage: Unified, Versatile, Highly Dynamic Many-to-many Image Generation introduces a versatile image generation technique capable of creating complex montages from multiple input images. This technology has applications in graphic design, advertising, and visual storytelling. Many-to-many image generation is a powerful tool for creative expression.
Advancements in Video Understanding
Video understanding remains a central focus in AI research, with applications ranging from surveillance to autonomous driving. This section showcases the latest papers addressing the challenges of analyzing and interpreting video content.
As mentioned earlier, GVLM: Geometry Grounded Vision Language Model with Unified 3D Reconstruction and Spatial Reasoning is also relevant in the context of video understanding, as its ability to integrate 3D reconstruction and spatial reasoning can enhance video analysis. The model's code is accessible on Github, allowing for further exploration and development in this area.
TimeViper: A Hybrid Mamba-Transformer Vision-Language Model for Efficient Long Video Understanding presents a novel hybrid architecture that combines Mamba and Transformer networks for efficient processing of long videos. The project page, available at https://xuboshen.github.io/TimeViper, provides additional details, and the code is available on Github. This research addresses the challenge of handling the temporal complexity of videos, which is crucial for many applications.
The Qwen3-VL Technical Report, spanning 42 pages, provides a comprehensive overview of the Qwen3-VL model, a powerful vision-language model capable of processing both images and videos. Technical reports like this are valuable resources for researchers and practitioners, offering in-depth information about the model's architecture, training, and performance.
Force Prompting: Video Generation Models Can Learn and Generalize Physics-based Control Signals demonstrates that video generation models can learn and generalize physics-based control signals through a technique called Force Prompting. This camera-ready version, presented at NeurIPS 2025, includes code and interactive demos available at https://force-prompting.github.io/. This research opens up new possibilities for creating realistic and controllable videos.
Thinking With Bounding Boxes: Enhancing Spatio-Temporal Video Grounding via Reinforcement Fine-Tuning explores the use of reinforcement fine-tuning to improve spatio-temporal video grounding, which involves localizing objects and events in videos. This approach leverages bounding boxes to provide spatial information, enhancing the model's understanding of video content. Reinforcement learning is a powerful tool for fine-tuning AI models for specific tasks.
SocialNav: Training Human-Inspired Foundation Model for Socially-Aware Embodied Navigation focuses on training a foundation model for socially-aware embodied navigation, which is crucial for robots operating in human environments. This research aims to develop robots that can navigate safely and effectively while interacting with people. Socially-aware navigation is a critical capability for robots in real-world settings.
As previously mentioned, CtrlVDiff: Controllable Video Generation via Unified Multimodal Video Diffusion, with its 27 pages, 18 figures, and 9 tables, also contributes to video understanding by enabling controllable video generation. The project page, accessible at https://tele-ai.github.io/CtrlVDiff/, provides more information on this approach.
Pistachio: Towards Synthetic, Balanced, and Long-Form Video Anomaly Benchmarks introduces a new benchmark for video anomaly detection, focusing on synthetic, balanced, and long-form videos. Benchmarks like Pistachio are essential for evaluating and comparing different anomaly detection algorithms. Robust anomaly detection is crucial for security and safety applications.
Video-R4: Reinforcing Text-Rich Video Reasoning with Visual Rumination presents a method for reinforcing text-rich video reasoning through visual rumination, which involves iteratively processing visual information to improve understanding. This approach is particularly relevant for videos with significant textual content, such as instructional videos or news reports.
Benchmarking the Trustworthiness in Multimodal LLMs for Video Understanding addresses the important issue of trustworthiness in multimodal large language models for video understanding. This research evaluates the reliability and safety of these models, which is crucial for their deployment in real-world applications. Trustworthiness is a key factor in the adoption of AI systems.
ReasonAct: Progressive Training for Fine-Grained Video Reasoning in Small Models focuses on training small models for fine-grained video reasoning using a progressive training approach. This research aims to develop efficient video understanding systems that can run on resource-constrained devices. Efficient models are essential for applications like mobile video analysis.
A review on data fusion in multimodal learning analytics and educational data mining provides a comprehensive overview of data fusion techniques in multimodal learning analytics and educational data mining. This review is valuable for researchers and practitioners interested in using AI to improve education. Multimodal data analysis can provide richer insights into learning processes.
Unsupervised Memorability Modeling from Tip-of-the-Tongue Retrieval Queries, accepted at WACV 2026, explores the use of tip-of-the-tongue retrieval queries to model video memorability in an unsupervised manner. This research contributes to the understanding of how videos are remembered and forgotten. Memorability modeling has applications in areas like advertising and content recommendation.
**[LongVT: Incentivizing
