🔥 PaddleOCR-VL-1.6: Expanding the Frontier of Document Parsing with Under-Optimized Region Refinement and Progressive Post-Training

💡 The paper presents PaddleOCR-VL-1.6, an upgraded document parsing model that improves upon its predecessor PaddleOCR-VL-1.5. The problem with the previous model was that its errors were concentrated in under-optimized regions where the model was unstable, data coverage was sparse, or supervision was unreliable. To address this issue, the authors introduced a region-aware data optimization framework that identifies weak regions in the previous model and applies targeted enhancements to these regions, improving the reliability of supervision signals.

The method used in PaddleOCR-VL-1.6 involves a progressive post-training recipe based on curated data selection and reinforcement learning, which pushes the model's performance to a higher level through staged optimization. This approach allows for more efficient use of data and computational resources, rather than simply expanding the training corpus.

The results show that PaddleOCR-VL-1.6 achieves a state-of-the-art score of 96.33% on the OmniDocBench v1.6 benchmark, demonstrating strong competitiveness against top-tier visual language models. The paper provides a practical post-training recipe for the PaddleOCR-VL series, which can be used to further improve the performance of document parsing models. Overall, the contributions of the paper include a region-aware data optimization framework, a progressive post-training recipe, and a new state-of-the-art model for document parsing.

📅 Published on Jun 2

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.03264
• PDF: https://arxiv.org/pdf/2606.03264
• Project Page: https://www.paddleocr.com

🤖 Models citing this paper:
• https://huggingface.co/PaddlePaddle/PaddleOCR-VL-1.6
• https://huggingface.co/PaddlePaddle/PaddleOCR-VL-1.6-GGUF

🚀 Spaces citing this paper:
• https://huggingface.co/spaces/PaddlePaddle/PaddleOCR-VL-1.6_Online_Demo

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#DocumentParsing #RegionRefinement #PostTrainingTechniques #OpticalCharacterRecognition #PaddleOCR

🔥 WavTTS: Towards High-Quality Zero-Shot TTS via Direct Raw Waveform Modeling

💡 The paper introduces WavTTS, a novel text-to-speech model that directly generates raw waveforms, addressing the limitations of existing latent-space diffusion models. Current state-of-the-art models operate on compressed representations such as mel-spectrograms or VAE latents, which leads to information loss and non-end-to-end training. Directly modeling raw waveforms is theoretically beneficial but has been underexplored due to the long sequence length of audio signals.

To overcome this challenge, WavTTS employs a flow matching approach with a Diffusion Transformer architecture and a simple patchification strategy to directly model speech waveforms. The model also incorporates multi-scale mel-spectrogram supervision to provide perceptual guidance during training. Additionally, the authors investigate the impact of prediction targets and noise scheduling in waveform diffusion and develop an effective schedule design to improve generation quality.

The results show that WavTTS significantly narrows the performance gap with latent-space generative models and outperforms previous end-to-end speech generation models. Evaluations on open-source benchmarks demonstrate that WavTTS closely approaches the performance of current state-of-the-art latent generative zero-shot text-to-speech models. The findings of this paper demonstrate the feasibility of scaling diffusion-based text-to-speech models directly in the waveform space, opening a new direction for end-to-end speech generation.

📅 Published on Jun 2

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.03455
• PDF: https://arxiv.org/pdf/2606.03455
• Project Page: https://wavtts.github.io/

🤖 Models citing this paper:
• https://huggingface.co/worstchan/WavTTS
• https://huggingface.co/drbaph/WavTTS

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#TextToSpeechSynthesis #RawWaveformModeling #DiffusionTransformer #ZeroShotTTS #SpeechSynthesisTechniques

🚀 Level up your AI & Data Science skills with HelloEncyclo — a growing all-in-one platform featuring hands-on courses in LLMs, Deep Learning, MLOps, Data Engineering, and more.
✅ 13 courses live + 40+ coming soon
🎯 One access, lifetime updates
🔑 Use code: PRESALE-BOOK-WAVE-2GFG
👉 https://helloencyclo.com/?ref=HUSSEINSHEIKHO

🔥 Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models

💡 The paper introduces Ultralytics YOLO26, a unified real-time vision model family that addresses the limitations of existing YOLO detectors. The problem with current YOLO detectors is that they rely on non-maximum suppression at inference, have heavy detection heads due to Distribution Focal Loss, require long training schedules, and often fail to assign positive labels to small objects. To overcome these limitations, YOLO26 uses a dual-head design for native NMS-free end-to-end inference and removes Distribution Focal Loss entirely, resulting in a lighter head with unconstrained regression range.

The method involves a coordinated architecture and training advances, including the use of MuSGD, a hybrid Muon-SGD optimizer, Progressive Loss, which shifts supervision toward the inference-time head, and STAL, a label assignment strategy that guarantees positive coverage for small objects. The model also introduces task-specific head and loss designs for instance segmentation, pose estimation, and oriented detection, allowing it to produce consistent gains across tasks and scales.

The YOLO26 family spans five scales and supports detection, instance segmentation, pose estimation, classification, and oriented detection in a single pipeline, with an open-vocabulary extension, YOLO E-26, for text-, visual-, and prompt-free inference. The results show that YOLO26 achieves 40.9-57.5 mAP on COCO at 1.7-11.8 ms T4 TensorRT latency, advancing the accuracy-latency Pareto front over prior real-time detectors. Additionally, YOLO E-26x reaches 40.6 AP on LVIS minival under text prompting, demonstrating the effectiveness of the proposed approach. The code and models are available for further research and development.

📅 Published on Jun 2

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.03748
• PDF: https://arxiv.org/pdf/2606.03748
• Project Page: https://docs.ultralytics.com/models/yolo26

🤖 Models citing this paper:
• https://huggingface.co/Ultralytics/YOLO26

🚀 Spaces citing this paper:
• https://huggingface.co/spaces/Ultralytics/YOLO26
• https://huggingface.co/spaces/atalaydenknalbant/Yolo26
• https://huggingface.co/spaces/itsMarco-G/reachy_phone_home

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#YOLO26 #RealTimeVisionModels #EndToEndInference #ObjectDetectionAlgorithms #UnifiedVisionModels

Join our livestream with Marina Wyss, Senior Applied Scientist at Twitch, as we discuss how to break into AI Engineering in 2026.

Sign up for FREE and save your seat here: luma.com/qgz4g4r7

Why should you join?

Many people interested in AI Engineering are asking the same questions:

❓ Where do I start?
🤔 Do I need deep math first?
🧠 Should I focus on ML, LLMs, RAG, or AI agents?
🧭 How do I avoid wasting time learning the wrong things?
🚀 How do I go from learning to becoming hireable?

If you’re interested in AI Engineering but unsure how to approach it, this livestream is for you.

What you’ll learn

✦ What AI Engineering really is
✦ Where beginners should start
✦ What skills and topics actually matter
✦ Common mistakes to avoid
✦ Self-study vs bootcamp vs MSc
✦ How to think about becoming hireable in AI
✦ Practical advice from someone already working in the field

Sign up for FREE and save your seat: luma.com/qgz4g4r7

🔥 Cosmos 3: Omnimodal World Models for Physical AI

💡 The paper introduces Cosmos 3, an omnimodal world model that can process and generate multiple data types, including language, image, video, audio, and action sequences, through a unified mixture-of-transformers architecture. This model is designed to jointly process and generate different modalities, effectively combining vision-language models, video generators, world simulators, and world-action models into a single framework. The Cosmos 3 model achieves state-of-the-art performance in various understanding and generation tasks, demonstrating its ability to serve as a scalable and general-purpose backbone for embodied agents.

The key contribution of the paper is the development of a unified architecture that can handle multiple input-output configurations, allowing for seamless integration of different modalities. The model is evaluated on a diverse suite of tasks, including text-to-image and image-to-video generation, and policy modeling, and achieves superior performance compared to existing models. The authors also make their code, model checkpoints, curated synthetic datasets, and evaluation benchmark available under an open-source license to accelerate research and deployment in Physical AI.

The results show that the Cosmos 3 model establishes a new state-of-the-art in various tasks, and its post-trained models were ranked as the best open-source models in text-to-image and image-to-video generation, as well as policy modeling. The availability of the code and model checkpoints is expected to facilitate further research and development in Physical AI, and the Cosmos 3 model has the potential to become a widely-used backbone for embodied agents. Overall, the paper presents a significant contribution to the field of Physical AI, demonstrating the effectiveness of omnimodal world models in achieving state-of-the-art performance in a wide range of tasks.

📅 Published on Jun 1

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.02800
• PDF: https://arxiv.org/pdf/2606.02800
• Project Page: https://research.nvidia.com/labs/cosmos-lab/cosmos3/

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#OmnimodalLearning #MultimodalGeneration #PhysicalAI #WorldModels #EmbodiedIntelligence

🔥 GRAIL: Generating Humanoid Loco-Manipulation from 3D Assets and Video Priors

💡 The paper presents GRAIL, a digital generation pipeline that produces diverse humanoid manipulation and locomotion data through 3D asset composition and video foundation models. The problem addressed is the difficulty of scaling teleoperation and motion capture for robot control due to the need for physical setups, instrumented actors, and robot operation. GRAIL solves this by remaining fully virtual until deployment, composing 3D assets, simulator-ready scenes, and priors from video foundation models to synthesize interactions without rebuilding physical environments or teleoperating the robot.

The method involves starting from fully specified 3D configurations where object geometry, camera parameters, and other factors are known before video generation and reused during reconstruction. This setup allows for model-based object tracking, human motion estimation, and interaction-aware optimization to reconstruct metric 4D human-object interaction trajectories with reduced depth ambiguity and morphology mismatch. The recovered motions are then retargeted to a humanoid robot and used to train task-general trackers.

The results show that GRAIL produces over 20,000 sequences of various tasks such as pick-up, object manipulation, sitting, and terrain traversal. Using only GRAIL-generated data, the authors train egocentric visual policies through a sim-to-real pipeline and deploy them on a Unitree G1 humanoid robot, achieving 84 percent real-world success on diverse object pick-up and 90 percent success on stair-climbing. The paper demonstrates the effectiveness of GRAIL in generating realistic and diverse data for robot control, enabling successful sim-to-real transfer and deployment on a real robot.

📅 Published on Jun 3

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.05160
• PDF: https://arxiv.org/pdf/2606.05160
• Project Page: https://research.nvidia.com/labs/dair/grail/

📊 Datasets citing this paper:
• https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-Locomanipulation-GRAIL

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#HumanoidRobotics #LocoManipulation #Teleoperation #MotionCapture #RobotControl

🔥 Echo-Infinity: Learning Evolving Memory for Real-Time Infinite Video Generation

💡 The paper introduces Echo Infinity, a framework for real-time infinite video generation that addresses the limitations of existing autoregressive methods. The main problem with current methods is that they use predefined memory curation techniques, such as fixed-ratio compression or inference-time adaptation, which lead to loss of historical information and compounding errors. To overcome this, Echo Infinity employs a learnable evolving memory that dynamically filters, abstracts, and compresses any-length history at constant cost. This is achieved through learnable Memory Queries that are updated by attention and a gating mechanism when past frames are evicted from the local window. The Memory Queries are optimized end-to-end with video diffusion transformers, forming an evolving memory that supports arbitrary compression ratios with constant computation independent of video length. Additionally, the paper introduces the Unified Relative RoPE Recipe, which frees the model from the finite RoPE constraint and closes the train-test RoPE extrapolation gap. The results show that Echo Infinity achieves state-of-the-art performance in both long and short video generation, and demonstrates promising real-time rollouts for up to 24 hours, suggesting a practical path toward infinite video generation. Overall, the paper presents a novel approach to real-time infinite video generation that overcomes the limitations of existing methods and achieves impressive results.

📅 Published on Jun 3

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.04527
• PDF: https://arxiv.org/pdf/2606.04527
• Project Page: https://echo-team-joy-future-academy-jd.github.io/Echo-Infinity/

🤖 Models citing this paper:
• https://huggingface.co/Echo-Team/Echo-Infinity

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#InfiniteVideoGeneration #RealTimeVideoSynthesis #EvolvingMemoryModels #AutoregressiveVideoMethods #LearnableMemoryMechanisms

🔥 OVO-S-Bench: A Hierarchical Benchmark for Streaming Spatial Intelligence in Multimodal LLMs

💡 The paper introduces OVO-S-Bench, a comprehensive benchmark for evaluating the ability of multimodal language models to understand spatial information from continuous video streams. The problem addressed is that existing benchmarks for spatial intelligence either evaluate models on full videos or focus on events rather than spatial structure, and do not account for the need to reason about places and layouts from partial information.

To address this, the authors created a benchmark consisting of 1680 human-annotated questions spanning 348 source videos, with each question having a query timestamp and an evidence interval. The questions cover four levels of abstraction, from basic perception to complex spatial reasoning and mapping. The annotation process involved 12 trained annotators who also served as cross-reviewers, ensuring high quality through multiple rounds of review.

The results show that even the best performing model, Gemini-3.1-Pro, trails human experts by 27 points, with the most challenging task being allocentric mapping. Interestingly, models that are specifically fine-tuned for streaming and spatial tasks actually perform worse than their original backbones, suggesting that these models may not be effectively using the spatial information in the video streams. The authors also found that using chain-of-thought reasoning can amplify spatial errors when the model is not grounded in the stream.

Overall, the OVO-S-Bench benchmark provides a challenging testbed for evaluating and improving the spatial intelligence of multimodal language models, and highlights the need for further research in this area to address the limitations of current models.

📅 Published on Jun 2

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.03890
• PDF: https://arxiv.org/pdf/2606.03890
• Project Page: https://internlm.github.io/OVO-S-Bench/

━━━━━━━━━━━━━━━━━━━━━━━━
📢 By: https://xn--r1a.website/PaperNexus

#MultimodalLanguageModels #SpatialIntelligence #StreamingVideoAnalysis #VideoUnderstandingBenchmarks #MultimodalLLMEvaluation