Friday, July 10, 2026
Vidu S1 achieves real-time interactive video generation at 42 FPS with voice control; agent skills ecosystem consolidation continues with skills, superpowers, and governance repos dominating GitHub trending; linear attention architectures get systematic comparative study
Executive Summary
Thursday's AI landscape centers on a breakthrough in interactive video generation. Vidu S1 (108 upvotes) introduces real-time voice-controlled video generation at 540p/42fps on consumer GPUs, combining TurboDiffusion with TurboServe for infinite-length generation without quality degradation. This sits alongside CineMobile, which brings cinematic camera motion generation to mobile devices via efficient diffusion transformers, and OpenCoF, which repurposes video generation models as reasoning engines through Chain-of-Frame reasoning. The video generation stack is rapidly moving from research demos to deployment-ready systems.
The agent ecosystem continues its infrastructure moment from Wednesday. GitHub trending is again dominated by agent-skill repos: superpowers (+1,013 stars), agent-skills (+1,116 stars), and skills (+1,712 stars) consolidate the skill-sharing paradigm, while Microsoft's Agent Governance Toolkit addresses OWASP Agentic Top 10 security concerns and TencentDB Agent Memory tackles the long-term memory gap. On the research side, UniClawBench proposes the first universal benchmark for proactive agents, and Remember When It Matters introduces proactive memory agents that actively intervene rather than passively retrieve, addressing behavioral state decay in long-horizon tasks.
The model landscape sees continued diversification: GLM-5.2 leads with 3,786 likes, Qwen 3.6 variants proliferate across quantization formats with uncensored and aggressive fine-tunes, NVIDIA LocateAnything-3B dominates downloads at 1.4M for vision grounding, and Tencent Hy3 enters trending as a new MoE text generation model. Mistral's Leanstral-1.5-119B-A6B and DeepSeek-V4-Flash continue the efficiency-focused large model trend.
Researcher Notes
Vidu S1 represents a deployment inflection for interactive video generation. The 42 FPS on consumer GPUs is the real story—prior interactive video systems required data-center hardware, making them impractical for consumer applications. The combination of voice control with real-time generation opens a product category (live AI-controlled avatars, interactive storytelling, real-time game cinematics) that was previously gated on hardware costs. The fact that it handles infinite-length generation without drift or distortion addresses the temporal consistency problem that has plagued long-form video generation.
Video-Oasis's evaluation rethink exposes a measurement crisis. With 43 upvotes, the paper's critique resonates: current Video-LLM benchmarks conflate visual perception, linguistic reasoning, and knowledge priors, making it impossible to attribute performance gains to actual video understanding. This is the same measurement problem that plagued NLP benchmarks (GLUE/SuperGLUE) before the community developed probing methodologies. Expect this to catalyze a wave of diagnostic benchmarks that decompose video understanding into testable sub-capabilities.
The agent skills ecosystem is forming faster than expected. For the second consecutive day, agent-skill repos dominate GitHub trending. The key signal today is Microsoft's Agent Governance Toolkit covering all 10 OWASP Agentic Top 10 risks—this is security tooling arriving in parallel with the capability explosion, not lagging behind. Combined with TencentDB's agent memory solution and UniClawBench's universal agent benchmark, the ecosystem is simultaneously building capability (skills), infrastructure (memory, templates), evaluation (benchmarks), and security (governance). This coordinated maturation across all four pillars is rare and suggests genuine enterprise readiness rather than hype-driven adoption.
Linear attention's comparative study is a consolidation signal. The systematic comparison of DeltaNet, Gated DeltaNet, Kimi Delta Attention, and Gated DeltaNet-2 in a unified notation suggests the linear attention space is mature enough for head-to-head evaluation rather than isolated proposal papers. The cross-layer routing extension hints at the next frontier: attention mechanisms that dynamically route tokens to different processing strategies based on their position and content.
IdeaGene-Bench is a sleeper for AI-for-science. Representing scientific ideas as inheritable genomes and benchmarking lineage reasoning is a conceptually novel framing that could reshape how we evaluate AI scientific reasoning. Rather than testing whether an AI can answer questions about papers, it tests whether it can trace the intellectual DNA of ideas—a capability closer to how human researchers actually think about their field.
Themes & Trends
Real-Time Video Generation Goes Consumer-Grade
risingMultiple papers push video generation toward real-time, consumer-hardware deployment: Vidu S1 achieves 42fps on consumer GPUs, CineMobile targets mobile devices, and OpenCoF repurposes generation as a reasoning primitive.
Agent Skills Ecosystem Infrastructure
risingGitHub trending continues to be dominated by agent skill repositories, governance toolkits, and memory systems, signaling the maturation of agent infrastructure from ad-hoc scripts to composable, auditable, enterprise-ready components.
Evaluation Methodology Rethink
risingPapers challenge existing benchmark paradigms: Video-Oasis argues current video benchmarks conflate perception with reasoning, UniClawBench proposes capability-decomposed agent evaluation, and IdeaGene-Bench tests lineage reasoning rather than summarization.
Long-Context and Efficient Attention
stableJet-Long's dynamic bifocal RoPE and the linear attention comparative study reflect continued investment in making transformers efficient at long sequences, driven by agentic and RAG deployment requirements.
AI for Scientific Discovery
risingDrugGen-2 introduces disease-aware molecular generation and IdeaGene-Bench benchmarks scientific lineage reasoning, showing AI-for-science expanding from tool augmentation to conceptual reasoning about scientific inheritance.
Trending Papers (14)
Vidu S1: A Real-Time Interactive Video Generation Model
High RelevanceJintao Zhang, Kai Jiang, Jintao Chen, Xu Wang, Yang Luo et al. — Tsinghua University, Shengshu Technology
Introduces Vidu S1, a real-time interactive video generation model supporting voice control of digital characters with infinite-length generation at 540p/42fps on consumer GPUs, using TurboDiffusion and TurboServe architectures.
Key Findings
- •
Achieves real-time video generation at up to 42 FPS on consumer GPUs via TurboDiffusion and TurboServe
- •
Supports infinite-length generation without blurring, drift, or visual distortion
- •
Enables voice-controlled digital characters with customizable images of real people, anime, and pets
Video-Oasis: Rethinking Evaluation of Video Understanding
High RelevanceGeuntaek Lim, Sungjune Park, Jaeyun Lee, Inwoong Lee, Taeoh Kim — NAVER AI Lab, KAIST
Re-examines evaluation criteria for video understanding, arguing that current benchmarks conflate visual perception, linguistic reasoning, and knowledge priors, making it difficult to attribute Video-LLM performance to actual video comprehension.
Key Findings
- •
Current Video-LLM benchmarks fail to isolate visual perception from linguistic reasoning and knowledge priors
- •
Proposes shared evaluation criteria that have been overlooked despite proliferation of high-level reasoning benchmarks
- •
Demonstrates that many benchmark gains stem from language priors rather than genuine video understanding
Why Can't I Open My Drawer? Mitigating Object-Driven Shortcuts in Zero-Shot Compositional Action Recognition
High RelevanceGeo Ahn, Inwoong Lee, Taeoh Kim, Minho Shim, Dongyoon Wee — NAVER AI Lab, Seoul National University
Tackles a key failure mode in zero-shot compositional action recognition where models predict verbs via object-driven shortcuts rather than temporal evidence, proposing diagnostic metrics and mitigation strategies.
Key Findings
- •
Identifies that models rely on object class labels rather than temporal motion evidence to predict verbs
- •
Sparse compositional supervision and verb-object learning asymmetry promote shortcut learning
- •
Proposed diagnostic metrics reveal the extent of object-driven shortcuts across existing methods
UniClawBench: A Universal Benchmark for Proactive Agents on Real-World Tasks
High RelevanceZhekai Chen, Chengqi Duan, Kaiyue Sun, Bohao Li, Yuqing Wang — Peking University, Beijing Academy of Artificial Intelligence
Proposes a universal benchmark for proactive agents that addresses limitations of existing sandboxed, single-turn evaluation paradigms by introducing a capability-decomposed task taxonomy for real-world agent assessment.
Key Findings
- •
Existing benchmarks mix multiple model capabilities within the same task category, obscuring strengths and weaknesses
- •
Introduces capability-decomposed taxonomy that separates planning, tool use, and interaction capabilities
- •
Supports multi-turn evaluation in real-world environments rather than sandboxed simulations
Ideas Have Genomes: Benchmarking Scientific Lineage Reasoning and Lineage-Grounded Idea Generation
High RelevanceYifan Zhou, Qihao Yang, Yan Li, Donggang Li, Xiru Hu — University of Illinois Urbana-Champaign, Tsinghua University
Presents IdeaGene-Bench, a benchmark for scientific lineage reasoning that represents papers as minimal inheritable units, testing whether AI can trace the intellectual inheritance structure of scientific ideas.
Key Findings
- •
Represents scientific ideas as genomes with inheritable mechanisms, limitations, and recombinations
- •
Benchmarks both lineage reasoning (tracing idea inheritance) and lineage-grounded idea generation
- •
Reveals that current LLMs struggle with fine-grained lineage attribution despite strong summarization
LongE2V: Long-Horizon Event-based Video Reconstruction, Prediction, and Frame Interpolation with Video Diffusion Models
Cheng-De Fan, Chun-Wei Tuan Mu, Chen-Wei Chang, Chin-Yang Lin, Kun-Ru Wu — National Tsing Hua University, National Yang Ming Chiao Tung University
Proposes LongE2V, leveraging pre-trained video diffusion priors for event-based video reconstruction, prediction, and frame interpolation, achieving long-term stability through Autoregressive Rolling Overlap and Cross-Clip Attention.
Key Findings
- •
Leverages pre-trained video diffusion priors for event-based tasks, achieving high data efficiency
- •
Autoregressive Rolling Overlap enables long-horizon temporal consistency
- •
Cross-Clip Attention maintains coherence across long sequences while preserving perceptual quality
Enhancing In-context Panoramic Generation via Geometric-aware Pretraining
Haoran Feng, Ruiyang Zhang, Longyi Zhang, Dizhe Zhang, Lu Qi — HKUST, The University of Hong Kong
Presents Canvas360, a two-stage framework for in-context panoramic generation combining geometry-aware pretraining with task-specific fine-tuning, trained on Canvas360Dataset with 1M high-quality paired panoramic samples.
Key Findings
- •
Canvas360Dataset provides 1M high-quality paired panoramic samples across style transfer, inpainting, outpainting, and editing
- •
Geometry-aware pretraining addresses the lack of panoramic-specific training data
- •
Two-stage framework generalizes across diverse in-context panoramic generation scenarios
DrugGen 2: A disease-aware language model for enhancing drug discovery
High RelevanceAli Motahharynia, Mohammadreza Ghaffarzadeh-Esfahani, Mahsa Sheikholeslami, Navid Mazrouei, Matin Irajpour — Isfahan University of Medical Sciences, Isfahan University of Technology
Introduces DrugGen-2, a generative model that designs small molecules conditioned on both disease ontology and target protein sequences, addressing the gap where most drug design methods neglect disease context.
Key Findings
- •
First model to condition molecular generation on both disease ontology and target protein sequences
- •
Fine-tuned GPT-2 on curated disease-target-molecule triplets captures disease-specific therapeutic relationships
- •
Generates molecules with higher predicted binding affinity and drug-likeness compared to disease-agnostic baselines
Jet-Long: Efficient Long-Context Extension with Dynamic Bifocal RoPE
High RelevanceHaozhan Tang, Zerui Wang, Yuxian Gu, Song Han, Han Cai — MIT, NVIDIA
Proposes Dynamic Bifocal RoPE for zero-shot context extension that dynamically adjusts rescaling factors to preserve short-context fidelity while extending to long contexts, addressing the fixed-factor tradeoff in existing methods.
Key Findings
- •
Dynamic bifocal rescaling preserves short-context performance while extending context windows by 10x+
- •
Zero-shot approach requires no fine-tuning, making it immediately deployable on existing checkpoints
- •
Outperforms fixed-factor methods (YaRN, NTK-aware) on both short and long context benchmarks
CineMobile: On-Device Image-to-Video Diffusion for Cinematic Camera Motion Generation
Xuyao Huang, Zelai Deng, Xu Wang, Xizhong Xiao, Zhijie Deng — Shanghai Jiao Tong University, Kuaishou Technology
Proposes CineMobile for efficient on-device image-to-video generation targeting cinematic camera motion effects, bridging the gap between DiT video generation quality and mobile deployment constraints.
Key Findings
- •
Achieves on-device video generation for cinematic effects like bullet time, dolly zoom, and slow motion
- •
Reduces computational overhead through architecture-level and inference-level optimizations for mobile DiTs
- •
Maintains visual quality comparable to full-scale DiT models despite significant parameter and compute reduction
OpenCoF: Learning to Reason Through Video Generation
High RelevanceXinyan Chen, Ziyu Guo, Renrui Zhang, Dongzhi Jiang, Hongsheng Li — The Chinese University of Hong Kong, Shanghai AI Laboratory
Proposes Chain-of-Frame (CoF) reasoning where reasoning unfolds through temporally connected video frames, introducing dedicated training data and architectural designs for video-based reasoning beyond Chain-of-Thought.
Key Findings
- •
Reasoning through temporally connected video frames offers a distinct path from text-based Chain-of-Thought
- •
Existing video generators lack diverse supervision and dedicated designs for CoF reasoning
- •
Purpose-built training and architecture significantly improve reasoning accuracy in visual-spatial domains
Linear Attention Architectures: Mechanisms, Trade-offs, and Cross-Layer Routing
High RelevanceTommaso Cerruti, Tim Rieder, George Rowlands, Lingfeng Jin, Imanol Schlag — IDSIA, University of Lugano, ETH Zurich
Provides the first systematic comparison of softmax attention against four recent linear-attention architectures (DeltaNet, Gated DeltaNet, Kimi Delta Attention, Gated DeltaNet-2) using a unified recurrent-memory notation.
Key Findings
- •
Unified recurrent-memory notation reveals how architectures differ in expressivity, memory decay, and write control
- •
Introduces cross-layer routing as an extension that dynamically routes tokens across different attention strategies
- •
Identifies specific trade-offs between computational efficiency and expressivity across the four architectures
Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents
High RelevanceYifan Wu, Lizhu Zhang, Yuhang Zhou, Mingyi Wang, Bo Peng — Microsoft Research, Peking University
Studies memory as an active intervention mechanism rather than passive retrieval for long-horizon agents, identifying 'behavioral state decay' as a key failure mode and proposing proactive memory agents to address it.
Key Findings
- •
Identifies 'behavioral state decay' as the failure mode where decision-relevant state gets buried as trajectories grow
- •
Proactive memory agents actively surface and intervene with relevant context rather than waiting for retrieval queries
- •
Demonstrates significant performance improvements on long-horizon tasks compared to passive memory approaches
UP: Unbounded Positive Asymmetric Optimization for Breaking the Exploration-Stability Dilemma
Chongyu Fan, Pengfei Liu, Jingjia Huang, Sijia Liu, Yi Lin — Shanghai Jiao Tong University, Michigan State University
Proposes unbounded positive asymmetric optimization to break the exploration-stability dilemma in importance-sampling-based RL for LLMs, allowing unrestricted positive updates while maintaining stability.
Key Findings
- •
Standard clipping mechanisms constrain policy updates too strictly, limiting exploration of beneficial trajectories
- •
Unbounded positive asymmetry allows large updates for high-quality trajectories while preventing catastrophic instability
- •
Achieves improved sample efficiency on complex reasoning tasks compared to PPO and GRPO baselines
Trending Models (11)
Zhipu AI (ZAI) · text-generation · MoE
Large-scale MoE language model with dense-sparse architecture supporting both English and Chinese, continuing the GLM series with significant architecture improvements.
NVIDIA · feature-extraction · 3B
Vision-language model for open-vocabulary object detection and grounding, built on the EAGLE architecture with 3B parameters, supporting text-based spatial queries.
HauhauCS · image-text-to-text · 35B-A3B
Uncensored and aggressively fine-tuned Qwen 3.6 MoE variant with 35B total parameters and 3B active, supporting vision and multimodal inputs in GGUF format.
Empero AI · image-text-to-text · 9B
GGUF-quantized 9B model with 1M context window supporting reasoning, function calling, multimodal vision, and specialized capabilities in cybersecurity and biomedical domains.
Baidu · image-text-to-text · unknown
Vision-language OCR model supporting unlimited-length document recognition across multiple languages, combining feature extraction with image-text understanding.
DeepReinforce AI · text-generation · 35B
35B parameter text generation model distributed in GGUF format, optimized for efficient inference with MIT license.
Tencent · text-generation · MoE
Tencent's Hunyuan V3 MoE model for text generation, part of the Hunyuan series with improved architecture and training.
InternScience · image-text-to-text · MoE
Agentic vision-language MoE model built on Qwen 3.5 architecture, designed specifically for agent tasks with tool use and visual understanding capabilities.
Google · tabular-regression · unknown
Foundation model for tabular data supporting zero-shot and in-context learning for both regression and classification tasks on structured data.
Community (yuxinlu1) · text-generation · 12B
GGUF-quantized Gemma 4 12B fine-tuned for agentic coding and terminal use with enhanced reasoning and tool-use capabilities.
BottleCap AI · image-text-to-text · 27B
Token-efficient thinking model based on Qwen 3.6 27B, optimized for efficient reasoning with reduced token overhead while maintaining quality.
Trending GitHub Repos (12)
Curated collection of production-grade engineering skills for AI coding agents, directly from a developer's .claude directory. Represents the growing skill-sharing ecosystem for agent capabilities.
Production-grade engineering skills for AI coding agents with a focus on best practices and robust implementations across common development tasks.
An agentic skills framework and software development methodology that works, providing structured approaches to AI-assisted development workflows.
MCP server for Claude providing terminal control, file system search, and diff-based file editing capabilities, enabling deeper agent-OS integration.
AI agent skill that researches any topic across Reddit, X, YouTube, Hacker News, Polymarket, and the web, then synthesizes a grounded summary of the last 30 days.
Reusable computer vision tools library for building detection, segmentation, and tracking pipelines with a clean API across multiple model backends.
Framework for building local voice agents with open-source models, enabling real-time speech-to-speech interaction on local hardware.
Fully local long-term memory for AI agents via a 4-tier progressive pipeline with zero external API dependencies, supporting persistent agent state management.
CLI tool for configuring and monitoring Claude Code instances with pre-built templates for common development workflows.
Fastest KV cache layer for LLM serving, enabling efficient prefix caching and context reuse for high-throughput inference.
AI Agent Governance Toolkit providing policy enforcement, zero-trust identity, execution sandboxing, and reliability engineering for autonomous agents, covering all 10 OWASP Agentic Top 10 risks.
Framework for building and running transparent, understandable, and trustworthy AI agents with strong observability and debugging capabilities.