Hugging Face Trending Papers

This episode explores groundbreaking AI research, featuring Helios, a real-time long video generation model; Proact-VL, a proactive VideoLLM for real-time AI companions; and T2S-Bench & Structure-of-Thought, a new benchmark and prompting technique for text-to-structure reasoning.### Featured Papers* **Helios: Real Real-Time Long Video Generation Model** * **Key Insight:** Helios is the first 14B video generation model capable of real-time (19.5 FPS) minute-scale video generation on a single H100 GPU, achieving high quality by addressing long-video drifting and optimizing for efficiency. * **Paper Link:** [https://arxiv.org/pdf/2603.04379.pdf](https://arxiv.org/pdf/2603.04379.pdf)* **Proact-VL: A Proactive VideoLLM for Real-Time AI Companions** * **Key Insight:** Proact-VL introduces a framework for creating proactive, real-time interactive AI companions, particularly for gaming scenarios like commentators and guides, by enabling low-latency inference and autonomous decision-making. * **Paper Link:** [https://arxiv.org/pdf/2603.03447.pdf](https://arxiv.org/pdf/2603.03447.pdf)* **T2S-Bench & Structure-of-Thought: Benchmarking and Prompting Comprehensive Text-to-Structure Reasoning** * **Key Insight:** This work introduces Structure-of-Thought, a prompting technique that guides models to construct intermediate text structures, and T2S-Bench, the first benchmark designed to evaluate and improve models' text-to-structure reasoning capabilities. * **Paper Link:** [https://arxiv.org/pdf/2603.03790.pdf](https://arxiv.org/pdf/2603.03790.pdf)

# Hugging Face Trending Papers Episode SummaryIn this episode, we discuss two trending papers, "Large-Scale Agentic RL for High-Performance CUDA Kernel Generation" and "Language-Agnostic SWE Task Collection at Scale". The first paper presents CUDA Agent, a large-scale reinforcement learning system that optimizes GPUs for deep learning, and the second introduces SWE-rebench V2, a language-agnostic, automated pipeline for collecting real-world software engineering tasks for training software engineering agents.## Papers Discussed- "Large-Scale Agentic RL for High-Performance CUDA Kernel Generation" introduces CUDA Agent, a system that fundamentally improves GPU optimization ability for deep learning using scalable data synthesis, skill-augmented CUDA development, and reinforcement learning techniques. The system achieves state-of-the-art results on KernelBench. [Read the paper](https://arxiv.org/pdf/2602.24286)- "Language-Agnostic SWE Task Collection at Scale" presents SWE-rebench V2, an automated pipeline for collecting real-world software engineering tasks and constructing reinforcement learning training environments at scale. The pipeline has constructed a dataset of 32,000+ tasks spanning 20 languages and 3,600+ repositories. [Read the paper](https://arxiv.org/pdf/2602.23866)## Additional Links- Project page for CUDA Agent: [https://cuda-agent.github.io/](https://cuda-agent.github.io/)Remember to follow or subscribe for the latest in AI research, and stay curious!

Kandinsky 5.0: A Family of Foundation Models for Image and Video Generation**Source:** huggingface_daily**URL:** https://huggingface.co/papers/2511.14993**Key Points:**- Problem: The research addresses the challenges in high-resolution image and video generation, particularly the scalability and computational complexity associa...- Method: The authors introduce Kandinsky 5.0, a family of foundation models comprising three core variants: Kandinsky 5.0 Image Lite, Kandinsky 5.0 Video Lite,...- Results: Kandinsky 5.0 achieves state-of-the-art performance in high-resolution image and 10-second video synthesis, demonstrating superior generation quality ...- Implications: Kandinsky 5.0 has significant implications for the research community by providing an open-source framework that advances the accessibility and develo...

# Episode SummaryIn this episode of Hugging Face Trending Papers, we delve into the latest AI research with three top trending papers from arXiv. We explore MiroThinker's interaction scaling for open-source research agents, the new paradigm of "Thinking with Video" for multimodal reasoning, and Lumine's approach to building generalist AI agents for 3D open-world environments.# Mentioned Papers1. ["MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling"](https://arxiv.org/pdf/2511.11793) - This paper presents MiroThinker, an open-source research agent that improves tool-augmented reasoning and information-seeking capabilities by focusing on efficient interaction scaling.2. ["Thinking with Video: Video Generation as a Promising Multimodal Reasoning Paradigm"](https://arxiv.org/pdf/2511.04570) - The authors propose "Thinking with Video," a new paradigm that uses video generation models to bridge visual and textual reasoning, overcoming limitations of current "Thinking with Text" and "Thinking with Images" paradigms.3. ["Lumine: An Open Recipe for Building Generalist Agents in 3D Open Worlds"](https://arxiv.org/pdf/2511.08892) - Lumine introduces a recipe for developing AI agents capable of completing complex missions in 3D open-world environments, demonstrating strong zero-shot cross-game generalization.

Papers discussed:1. [Scaling Latent Reasoning via Looped Language Models](https://arxiv.org/pdf/2510.25741): This paper introduces a new kind of pre-trained looped language models, Ouro, which improves reasoning capabilities by integrating reasoning into the pre-training phase. The models have demonstrated superior performance due to enhanced knowledge manipulation capabilities.2. [Concerto: Joint 2D-3D Self-Supervised Learning Emerges Spatial Representations](https://arxiv.org/pdf/2510.23607): The Concerto model combines 2D and 3D learning for improved spatial cognition in AI. This integration, involving 3D intra-modal self-distillation with 2D-3D cross-modal joint embedding, has yielded promising results in 3D scene perception and set new benchmarks in scene understanding.3. [RECODE: Unify Plan and Action for Universal Granularity Control](https://arxiv.org/pdf/2510.23564): RECODE is a new paradigm that unifies planning and action within a single code representation, facilitating dynamic control of decision granularity. This approach has proven effective in enhancing inference performance and training data efficiency.

**Episode Summary:**This episode dives into cutting-edge advancements for Large Language Models, covering new methods to enhance reasoning reliability and efficiency, and introducing lightweight memory systems for more effective long-term interaction.**Featured Papers:*** **A Theoretical Study on Bridging Internal Probability and Self-Consistency for LLM Reasoning** * *Key Insight:* Introduces RPC, a novel method that theoretically and empirically improves LLM reasoning by combining self-consistency and perplexity, achieving exponential error convergence and reducing sampling costs by 50%. * *Link:* https://arxiv.org/pdf/2510.15444* **LIGHTMEM: LIGHTWEIGHT AND EFFICIENT MEMORY-AUGMENTED GENERATION** * *Key Insight:* Presents LightMem, a human-memory-inspired system that enables LLMs to leverage historical interactions efficiently, significantly reducing token usage, API calls, and runtime while boosting accuracy. * *Link:* https://arxiv.org/pdf/2510.18866 * **DeepAnalyze: Agentic Large Language Models for Autonomous Data Science** * *Key Insight:* Introduces an agentic LLM framework for autonomous data science, automating the entire process from raw data to analyst-graded research reports using multi-agent collaboration and feedback reasoning. * *Link:* https://arxiv.org/pdf/2510.16872

In this episode of Hugging Face Trending Papers, we discuss three exciting AI research papers: "Less is More: Recursive Reasoning with Tiny Networks", "Agent Learning via Early Experience", and "Paper2Video: Automatic Video Generation from Scientific Papers". ## Papers Discussed1. **[Less is More: Recursive Reasoning with Tiny Networks](https://arxiv.org/pdf/2510.04871)**: This paper introduces a Tiny Recursive Model that significantly improves accuracy on hard question-answer problems, using a simpler recursive reasoning approach and beating Large Language Models on complex tasks. 2. **[Agent Learning via Early Experience](https://arxiv.org/pdf/2510.08558)**: This research paper presents a new paradigm called "early experience", where AI agents learn from their own actions. The approach improved effectiveness and out-of-domain generalization in diverse environments.3. **[Paper2Video: Automatic Video Generation from Scientific Papers](https://arxiv.org/pdf/2510.05096)**: This paper presents Paper2Video, a multi-agent framework designed to automate the labor-intensive process of generating academic presentation videos from scientific papers. ## Episode Links- [Paper 1: Less is More: Recursive Reasoning with Tiny Networks](https://arxiv.org/pdf/2510.04871)- [Paper 2: Agent Learning via Early Experience](https://arxiv.org/pdf/2510.08558)- [Paper 3: Paper2Video: Automatic Video Generation from Scientific Papers](https://arxiv.org/pdf/2510.05096)

In this episode, we discuss three trending AI research papers. We delve into the challenges and solutions related to code language models, video generation, and reinforcement learning. Key Points Discussed#LongCodeZip: Compress Long Context for Code Language Models- LongCodeZip is a novel framework for compressing code for Large Language Models (LLMs)- It addresses the issue of high API costs and generation latency associated with processing long inputs in codebases- The framework uses a dual-stage compression strategy, enabling it to preserve essential information while reducing context size- Evaluations show that LongCodeZip consistently outperforms baseline methods- This research could improve the efficiency and capability of code intelligence applications#Self-Forcing++: Towards Minute-Scale High-Quality Video Generation- The paper addresses the computational cost of generating long videos with diffusion models- It proposes an approach that uses teacher models to guide student models through sampled segments from self-generated long videos- This method allows for video length scaling up to 20× beyond the teacher's capability- The authors manage to generate videos up to 4 minutes and 15 seconds long, substantially outperforming baseline methods#EXGRPO: Learning to Reason from Experience- The paper investigates what makes a reasoning experience valuable in the context of Reinforcement Learning from Verifiable Rewards (RLVR)- The authors propose a framework that organizes and prioritizes valuable experiences- The approach aims to balance exploration with experience exploitation for efficient and scalable RLVR### Links to Papers- [LongCodeZip: Compress Long Context for Code Language Models](https://arxiv.org/pdf/2510.00446 )- [Self-Forcing++: Towards Minute-Scale High-Quality Video Generation](https://arxiv.org/pdf/2510.02283 )- [EXGRPO: Learning to Reason from Experience](https://arxiv.org/pdf/2510.02245 )

In today’s episode of Hugging Face Trending Papers, we explore three cutting-edge ideas reshaping how we interact with AI and research.First, we dive into Paper2Agent, a framework that transforms static research papers into interactive AI agents, making findings more transparent and usable. Next, we look at Scaling Agents via Continual Pre-training, which pushes the boundaries of agent reliability by teaching them through lifelong learning. Finally, we cover Documenting Machine Learning with AI Partners, a vision of AI tools that act as collaborative lab partners to capture, explain, and even co-author machine learning workflows.Listen in for benchmarks, methods, and real-world implications — and discover how these papers could change the way we do science.Papers discussed:Paper2Agent: Reimagining Research Papers As Interactive and Reliable AI AgentsScaling Agents via Continual Pre-trainingDocumenting Machine Learning with AI Partners

In this episode, we unpack three fresh arXiv papers shaping how AI creates, reasons, and acts. First, arXiv:2509.22622 explores real-time, steerable long-form video generation you can guide on the fly (PDF: https://arxiv.org/pdf/2509.22622). Next, arXiv:2509.25454 integrates tree search directly into reinforcement-learning training for verifiable reasoning—think math and code with checkable rewards (PDF: https://arxiv.org/pdf/2509.25454). Finally, arXiv:2510.01051 introduces a unified “gym” for multi-turn, tool-using LLM agents so results are comparable and scalable (PDF: https://arxiv.org/pdf/2510.01051). We break down why each matters, the key technical ideas, and what they could unlock for creators, engineers, and autonomous AI workflows.

1. RLAIF at Scale: Reinforcement Learning from AI Feedback for Multi-Turn ReasoningThis paper explores using AI-generated feedback instead of expensive human labels to train reasoning models. The authors show that Reinforcement Learning from AI Feedback (RLAIF) can match or even outperform models trained with limited human feedback, especially in multi-turn reasoning tasks.2. Learning to Forget: Dynamic Memory Compression in Long-Context TransformersThe authors propose a method for making transformers more efficient on long contexts by teaching them to “forget” unimportant details. Their dynamic memory compression reduces memory usage by over 40% while maintaining — and sometimes improving — accuracy on long-sequence benchmarks.3. VidAgent: Scalable Video Agents with Spatio-Temporal ReasoningThis work introduces VidAgent, a system that can understand and reason over long videos by grounding events in both space and time. It achieves state-of-the-art performance on video QA benchmarks and opens up possibilities for advanced video search and monitoring applications.

Today we cover three standout arXiv releases shaping vision, language, and evaluation. First, PANORAMA surveys the rise of omnidirectional, 360° perception for embodied AI—why standard pinhole vision isn’t enough, where datasets and models fall short, and how new backbones and adaptation methods are closing the gap. Read: https://arxiv.org/pdf/2509.12989 (arXiv:2509.12989).Next, the HALA technical report details an Arabic-centric instruction and translation pipeline—from FP8 translator teachers to multi-million sample corpora—powering models from 350M to 9B with strong benchmark gains. Read: https://arxiv.org/pdf/2509.14008 (arXiv:2509.14008).Finally, GenExam proposes a multidisciplinary “exam” for text-to-image models, revealing how strict, knowledge-heavy prompts expose major gaps in today’s generators. Read: https://arxiv.org/pdf/2509.14232 (arXiv:2509.14232).

In today’s 5–6 minute roundup, we cover: (1) SAPO’s decentralized RL that shares rollouts across a swarm for cheaper, faster LM post-training (arXiv:2509.08721 PDF), (2) VLA-Adapter’s “Bridge Attention” that makes small vision-language-action models both fast and state-of-the-art on robotics tasks (arXiv:2509.09372 PDF), and (3) HuMo’s unified generator coordinating text, reference images, and audio for people-centric video with strong identity + lip-sync (arXiv:2509.08519 PDF). Subscribe for crisp takes on what was done, why it matters, and where it might go next.

In this episode, we cover three HuggingFace trending AI papers shaping the future of alignment, training, and creativity.How models can reason over boundaries to stick to instructions (arXiv:2509.14760)How populations of models can evolve without labels through consensus and novelty (arXiv:2509.15194)How autoregressive generators can understand before they generate images (arXiv:2509.15185)Three different paths, one goal: building smarter, safer, and more creative AI.

Smarter computer agents, better reasoning, and robot manipulation breakthroughs.Today’s PapersScaleCUA: Scaling Open-Source Computer Use Agents with Cross-Platform Data🔗 arXiv:2509.15221➡️ Large dataset across 6 OSs and 3 task domains; closed-loop pipeline of auto-agents + human curation; big benchmark gains for GUI agents.FlowRL: Matching Reward Distributions for LLM Reasoning🔗 arXiv:2509.15207➡️ Shifts RL objective from reward maximization to reward distribution matching; preserves diverse reasoning paths; strong math & code benchmark results.RynnVLA-001: Using Human Demonstrations to Improve Robot Manipulation🔗 arXiv:2509.15212➡️ Two-stage pretraining on 12M ego-centric videos + trajectory-aware modeling; adds ActionVAE for action compression; stronger transfer to real robot tasks.