🤗 Daily Paper(2025-09-11)

[deep.di...@gmail.com]

🤗 Daily Paper Newsletter Hope you found some gems! This newsletter delivers you the curated list of papers by 🤗 Daily Papers. project page 🤗 daily paper 3D and 4D World Modeling: A Survey Published at 2025-09-04 #ML This review focuses on 3D and 4D world modeling, a crucial aspect of AI research that enables agents to comprehend and predict their dynamic surroundings. The study provides a detailed classification of various approaches, such as video-based, occupancy-based, and LiDAR-based methods, and offers a comprehensive overview of datasets and evaluation metrics for 3D/4D settings, all while identifying key challenges and potential research directions in the field.... Read More Hunyuan-MT Technical Report Published at 2025-09-05 #ML The report presents Hunyuan-MT-7B, a new open-source multilingual translation model that supports 33 languages, with a focus on Mandarin and minority languages. Additionally, it introduces Hunyuan-MT-Chimera-7B, a model that combines outputs from Hunyuan-MT-7B under different settings to improve performance, which outperforms other state-of-the-art models in various language translations.... Read More P3-SAM: Native 3D Part Segmentation Published at 2025-09-08 #ML The study presents P3-SAM, a model for fully automated segmentation of 3D objects into parts, which improves upon existing methods by providing precise results and strong robustness for complex objects. The model uses a feature extractor, multiple segmentation heads, and an IoU predictor, and is trained on a large dataset of 3.7 million models, achieving state-of-the-art performance.... Read More The Majority is not always right: RL training for solution aggregation Published at 2025-09-08 #ML This study presents a new method called AggLM that trains a model to improve its answers by combining multiple solutions, using rewards as feedback. The method works better than existing ones and can use solutions from various models, requiring fewer resources than other methods.... Read More A Survey of Reinforcement Learning for Large Reasoning Models Published at 2025-09-10 #ML This study reviews the progress of Reinforcement Learning in enhancing Large Language Models for complex reasoning tasks, and discusses the challenges and opportunities for further scaling towards Artificial SuperIntelligence.... Read More AgentGym-RL: Training LLM Agents for Long-Horizon Decision Making through Multi-Turn Reinforcement Learning Published at 2025-09-10 #ML The authors present AgentGym-RL, a new framework for training LLM agents in complex, real-world tasks using reinforcement learning, without relying on supervised fine-tuning. They also introduce ScalingInter-RL, a training approach that balances exploration and exploitation, allowing agents to develop diverse problem-solving strategies. The framework has been extensively tested and shown to perform well across various environments.... Read More RewardDance: Reward Scaling in Visual Generation Published at 2025-09-10 #ML The study presents RewardDance, a new framework for reward modeling in visual generation that addresses limitations in existing methods. By reformulating reward scores, RewardDance aligns with Vision-Language Models, enabling scaling of reward models and integration of task-specific instructions, reference examples, and chain-of-thought reasoning. Experiments show that RewardDance outperforms state-of-the-art methods and resolves the 'reward hacking' issue, resulting in diverse and high-quality ... Read More Published at Read More Tags are generated by Google's Gemini Pro API, and the summary and translation are generated by Upstage's SOLAR mini chat model derived from SOLAR-10.7B open LLM. (Experimental) The full paper is translated in korean with enko-t5-small-v0 model developed by Kim Kihyun. Visit Developer's Social Media