1. Anthropic releases nine Claude connectors for creative applications including Blender, Adobe, and Ableton

Anthropic has released nine connectors that allow Claude to integrate directly with creative software including Adobe Creative Cloud apps, Blender, Autodesk Fusion, Ableton Live, SketchUp, and others. The connectors enable users to control creative tools through natural language conversations with Claude, automating tasks like scene debugging in Blender, batch image processing in Adobe apps, and music production workflows in Ableton. Anthropic has also joined the Blender Development Fund as a patron to support ongoing development of the Python API that makes these integrations possible.

Read more: https://www.anthropic.com/news/claude-for-creative-work

Video Credit: @claudeai on X

2. Cursor launches SDK for building programmatic coding agents with same runtime as Cursor IDE

Cursor released its SDK in public beta, allowing developers to build programmatic coding agents using the same runtime, harness, and models that power Cursor’s desktop app, CLI, and web app. The TypeScript SDK enables agents to run locally, in Cursor’s cloud with dedicated VMs, or on self-hosted infrastructure, and includes features like codebase indexing, MCP server integration, skills, hooks, and subagents. The SDK is available via npm install and uses token-based consumption pricing.

Read more: https://cursor.com/changelog/sdk-release

Video Credit: @cursor_ai on X

3. ElevenLabs launches ElevenMusic platform for AI music creation, remixing and streaming

ElevenLabs has launched ElevenMusic, a revamped AI-powered music platform that allows users to create, remix, discover, and earn from music. The platform enables users to generate new songs from text prompts or remix existing tracks by changing genres or tempos. ElevenMusic is positioned as a fully licensed, artist-first service that includes a monetization model for creators to earn royalties. This represents ElevenLabs’ expansion beyond voice AI into comprehensive music generation and streaming services.

Read more: https://elevenlabs.io/blog/introducing-elevenmusic

Video Credit: @ElevenLabs on X

4. Google Gemini launches file generation feature for PDFs, Word, Excel, and Google Workspace documents

Google announced that Gemini can now generate downloadable files directly within chat conversations, including PDFs, Microsoft Word and Excel files, Google Docs, Sheets, and Slides. The feature supports additional formats like CSV, LaTeX, TXT, RTF, and Markdown, allowing users to create complete files from text prompts without manual copying and reformatting. The update is rolling out globally to all Gemini users and aims to streamline the workflow from brainstorming to finished documents.

Read more: https://blog.google/innovation-and-ai/products/gemini-app/generate-files-in-gemini/

Video Credit: @GeminiApp on X

That’s all for today’s Global AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

Keywords: RecursiveMAS, multi-agent systems, latent-space recursive computation, RecursiveLink module, gradient-based credit assignment

Category: AI Systems and Tools

Research Objective:

– Explore if agent collaboration can be scaled through recursion using the RecursiveMAS framework, extending recursive scaling principles from single models to multi-agent systems.

Research Methods:

– Developed a recursive multi-agent framework with a RecursiveLink module connecting agents and employing an inner-outer loop learning algorithm for system co-optimization through shared gradient-based credit assignment.

Research Conclusions:

– RecursiveMAS demonstrates superior efficiency and accuracy over traditional multi-agent systems, achieving an average accuracy improvement of 8.3%, 1.2 to 2.4 times faster inference speed, and a significant reduction in token usage by 34.6%-75.6% across multiple benchmarks.

Paper link: https://huggingface.co/papers/2604.25917

2. DV-World: Benchmarking Data Visualization Agents in Real-World Scenarios

Keywords: Data Visualization, Cross-Platform Evolution, Intent Alignment, Native Environment

Category: AI Systems and Tools

Research Objective:

– To introduce DV-World, a comprehensive benchmark designed to evaluate data visualization agents across professional lifecycles in the real world.

Research Methods:

– Developed a benchmark in three domains: DV-Sheet, DV-Evolution, and DV-Interact, incorporating Table-value Alignment and MLLM-as-a-Judge for evaluation.

Research Conclusions:

– State-of-the-art models perform below 50% in handling real-world data visualization tasks, highlighting significant challenges.

Paper link: https://huggingface.co/papers/2604.25914

3. Meta-CoT: Enhancing Granularity and Generalization in Image Editing

Keywords: Meta-CoT, Image Editing, Chain-of-Thought, Decomposability, Generalizability

Category: Computer Vision

Research Objective:

– To enhance image editing capabilities by decomposing editing operations into a task-target-understanding framework, improving granularity and generalization.

Research Methods:

– Proposes Meta-CoT, which breaks down single-image editing tasks into triplets and fundamental meta-tasks, incorporating a CoT-Editing Consistency Reward to align editing with Chain-of-Thought reasoning.

Research Conclusions:

– Meta-CoT achieves a 15.8% performance improvement across 21 editing tasks and demonstrates strong generalization to unseen tasks, with source code and benchmarks provided for public access.

Paper link: https://huggingface.co/papers/2604.24625

4. Mutual Forcing: Dual-Mode Self-Evolution for Fast Autoregressive Audio-Video Character Generation

Keywords: Mutual Forcing, autoregressive audio-video generation, joint audio-video modeling, self-distillation, training-inference consistency

Category: Generative Models

Research Objective:

– To efficiently generate audio-video content with long-horizon synchronization using a unified model that combines few-step and multi-step training modes.

Research Methods:

– Developed a framework called Mutual Forcing that integrates uni-modal generators into a combined model using a two-stage training strategy to optimize joint audio-video modeling and facilitate fast autoregressive generation.

Research Conclusions:

– Mutual Forcing eliminates the necessity for an additional teacher model, reduces overhead, and enhances model training with real paired data. It delivers competitive or superior results compared to existing methods with significantly fewer steps, demonstrating improved efficiency and quality.

Paper link: https://huggingface.co/papers/2604.25819

5. Co-Director: Agentic Generative Video Storytelling

Keywords: AI-generated summary, hierarchical multi-agent framework, semantic coherence, global optimization problem, multimodal self-refinement

Category: Generative Models

Research Objective:

– The study aims to enhance video storytelling by addressing it as a global optimization problem using a hierarchical multi-agent framework to maintain semantic coherence.

Research Methods:

– The framework utilizes a multi-armed bandit approach to guide creative direction and employs multimodal self-refinement loops to prevent identity drift and ensure sequence-level consistency.

Research Conclusions:

– Co-Director outperforms existing methods, offering a robust solution capable of generalizing across broader cinematic narratives and successfully evaluated using the GenAD-Bench dataset.

Paper link: https://huggingface.co/papers/2604.24842

6. Toward Scalable Terminal Task Synthesis via Skill Graphs

Keywords: SkillSynth, terminal task synthesis, scenario-mediated, skill graph, execution trajectories

Category: AI Systems and Tools

Research Objective:

– The objective of the study is to introduce SkillSynth, an automated framework aimed at enhancing the diversity and quality of execution trajectories for training terminal agents.

Research Methods:

– SkillSynth constructs a scenario-mediated skill graph to generate a wide array of terminal task instances, using multi-agent harnesses to instantiate these into executable tasks. The framework samples workflow paths from the graph, controlling the diversity of trajectories explicitly.

Research Conclusions:

– Experiments using Terminal-Bench show that SkillSynth effectively enhances the diversity of execution trajectories. Task instances generated by SkillSynth have been used to train the Hy3 Preview, improving its capabilities in terminal-based environments.

Paper link: https://huggingface.co/papers/2604.25727

7. MAIC-UI: Making Interactive Courseware with Generative UI

Keywords: Zero-code, Interactive STEM courseware, Incremental generation, Pedagogical rigor, Multi-modal understanding

Category: AI in Education

Research Objective:

– The objective of the study is to introduce MAIC-UI, a zero-code system designed to empower educators to create and edit interactive STEM courseware rapidly and efficiently, thus overcoming traditional barriers such as the need for HTML/CSS/JavaScript expertise.

Research Methods:

– The study employs a zero-code authoring system that utilizes structured knowledge analysis with multi-modal understanding, a generate-verify-optimize pipeline, and Click-to-Locate editing with Unified Diff-based incremental generation for rapid iteration cycles.

Research Conclusions:

– MAIC-UI decreases editing iterations and enhances learnability and controllability compared to direct Text-to-HTML generation. Classroom deployments demonstrated significant improvements in learning outcomes, fostering learning agency and reducing outcome disparities in students.

Paper link: https://huggingface.co/papers/2604.25806

8. IAM: Identity-Aware Human Motion and Shape Joint Generation

Keywords: identity-aware motion generation, body morphology, motion dynamics, multimodal signals, joint motion-shape generation

Category: Generative Models

Research Objective:

– Propose an identity-aware motion generation framework that models the relationship between body morphology and motion dynamics using multimodal signals.

Research Methods:

– Utilize multimodal signals such as natural language descriptions and visual cues for representing identity.

– Introduce a joint motion-shape generation paradigm to synthesize motion sequences alongside body shape parameters.

Research Conclusions:

– The framework improves motion realism and motion-identity consistency while maintaining high motion quality as demonstrated on motion capture datasets and large-scale in-the-wild videos.

Paper link: https://huggingface.co/papers/2604.25164

9. GoClick: Lightweight Element Grounding Model for Autonomous GUI Interaction

Keywords: GoClick, GUI element grounding, vision-language model, encoder-decoder architecture, Progressive Data Refinement

Category: Computer Vision

Research Objective:

– Introduce GoClick, a lightweight model for GUI element grounding on mobile devices, focusing on high accuracy and low computational needs.

Research Methods:

– Implement an encoder-decoder architecture to develop a model with only 230M parameters.

– Utilize Progressive Data Refinement techniques for data optimization, including task type filtering and data ratio adjustment.

Research Conclusions:

– GoClick achieves high visual grounding accuracy, comparable to larger models, while maintaining efficiency.

– Enhances GUI agent performance in a device-cloud collaboration framework by improving element localization and success rates.

Paper link: https://huggingface.co/papers/2604.23941

10. The Last Harness You’ll Ever Build

Keywords: AI agents, automated harness engineering, evolutionary loops, meta-learning

Category: AI Systems and Tools

Research Objective:

– The objective is to automate the deployment of AI agents by using a two-level framework that optimizes task-specific harnesses through evolutionary loops and meta-learning protocols.

Research Methods:

– The framework consists of two levels: the first level optimizes a worker agent’s harness for a single task using the Harness Evolution Loop, while the second level, the Meta-Evolution Loop, optimizes the evolution protocol to enable quick harness adaptation across diverse tasks.

Research Conclusions:

– The proposed framework shifts the need for manual harness engineering to an automated process, potentially eliminating the need for human intervention in adapting agents to new task domains.

Paper link: https://huggingface.co/papers/2604.21003

11. Preferences of a Voice-First Nation: Large-Scale Pairwise Evaluation and Preference Analysis for TTS in Indian Languages

Keywords: multilingual TTS, linguistic diversity, perceptual dimensions, SHAP analysis, Bradley-Terry modeling

Category: Natural Language Processing

Research Objective:

– The research aims to develop a controlled multidimensional pairwise evaluation framework for multilingual Text to Speech (TTS) systems, with a focus on linguistic control and perceptual annotation across 10 Indic languages.

Research Methods:

– The study involves evaluating 7 state-of-the-art TTS systems using over 5,000 native and code-mixed sentences, collecting more than 120,000 pairwise comparisons from 1,900 native raters. The evaluations are made across 6 perceptual dimensions including intelligibility, expressiveness, voice quality, liveliness, noise, and hallucinations, utilizing Bradley-Terry modeling and SHAP analysis.

Research Conclusions:

– The research constructs a multilingual leaderboard and interprets human preferences while analyzing the reliability of the leaderboard. It highlights the model strengths and trade-offs across different perceptual dimensions.

Paper link: https://huggingface.co/papers/2604.21481

12.

Paper link: 

13. Offline Evaluation Measures of Fairness in Recommender Systems

Keywords: recommender system fairness, fairness evaluation measures, AI Ethics and Fairness, robustness, guidelines

Category: AI Ethics and Fairness

Research Objective:

– The research aims to address the limitations in the evaluation of fairness in recommender systems by analyzing theoretical flaws and developing novel approaches to improve the robustness and applicability of fairness evaluation measures.

Research Methods:

– Conduct theoretical and empirical analysis of existing fairness evaluation measures to expose their limitations.

– Investigate a wide range of offline evaluation measures across different fairness notions, focusing on both users and items, and varying evaluation granularities.

Research Conclusions:

– Propose new evaluation approaches and measures that overcome existing limitations, thereby enhancing interpretability and applicability.

– Provide guidelines for selecting appropriate fairness evaluation measures, facilitating more precise application in practical scenarios and thus advancing the state-of-the-art in the offline evaluation of fairness in recommender systems.

Paper link: https://huggingface.co/papers/2604.25032

14. Seeing Isn’t Believing: Uncovering Blind Spots in Evaluator Vision-Language Models

Keywords: Vision-Language Models, Evaluator VLMs, image-to-text, text-to-image, perturbations

Category: Multi-Modal Learning

Research Objective:

– The study systematically evaluates the reliability issues of current Evaluator VLMs in detecting various types of output errors.

Research Methods:

– Introduces targeted perturbations across key error dimensions and evaluates 4 prominent VLMs using over 4000 perturbed instances and multiple evaluation techniques.

Research Conclusions:

– Current VLM evaluators have substantial blind spots, particularly with fine-grained compositional and spatial errors, revealing their unreliable nature for benchmarking and urging caution in their development use. Code and data have been made publicly available.

Paper link: https://huggingface.co/papers/2604.21523

15. AutoGUI-v2: A Comprehensive Multi-Modal GUI Functionality Understanding Benchmark

Keywords: AutoGUI-v2, autonomous agents, Vision-Language Models, functionality understanding, interaction logic

Category: Robotics and Autonomous Systems

Research Objective:

– The research aims to evaluate autonomous agents’ capability in understanding and predicting interactions within Graphical User Interfaces (GUIs) using the AutoGUI-v2 benchmark.

Research Methods:

– Utilizes a novel VLM-human collaborative pipeline for parsing multi-platform screenshots to create hierarchical functional regions and diverse evaluation tasks.

Research Conclusions:

– The evaluation found that open-source Vision-Language Models fine-tuned on agent data excel in functional grounding, whereas commercial models perform better in functionality captioning. However, all models exhibit challenges with complex interaction logic, indicating that deep functional understanding remains crucial for advancing GUI agents.

Paper link: https://huggingface.co/papers/2604.24441

16. A Systematic Post-Train Framework for Video Generation

Keywords: Supervised Fine-Tuning, Reinforcement Learning from Human Feedback, Group Relative Policy Optimization, temporal coherence, visual quality

Category: Generative Models

Research Objective:

– To bridge the gap between pretraining performance of video diffusion models and real-world deployment requirements by enhancing controllability, temporal coherence, and visual quality.

Research Methods:

– Four-stage post-training framework involving Supervised Fine-Tuning, Reinforcement Learning from Human Feedback with Group Relative Policy Optimization, Prompt Enhancement, and Inference Optimization.

Research Conclusions:

– The proposed pipeline significantly mitigates common artifacts and enhances controllability, visual aesthetics, while maintaining efficient sampling costs, offering a practical blueprint for real-world deployment.

Paper link: https://huggingface.co/papers/2604.25427

17. V-GRPO: Online Reinforcement Learning for Denoising Generative Models Is Easier than You Think

Keywords: Variational GRPO, ELBO-based surrogates, generative models, human preferences, text-to-image synthesis

Category: Generative Models

Research Objective:

– The primary goal is to improve text-to-image synthesis by aligning generative models more efficiently with human preferences using the Variational GRPO method.

Research Methods:

– This method combines ELBO-based surrogates with Group Relative Policy Optimization (GRPO), enhancing stability and efficiency in the alignment process.

Research Conclusions:

– Variational GRPO achieves state-of-the-art performance in text-to-image synthesis with significant speed improvements over previous methods like MixGRPO and DiffusionNFT.

Paper link: https://huggingface.co/papers/2604.23380

18. TCOD: Exploring Temporal Curriculum in On-Policy Distillation for Multi-turn Autonomous Agents

Keywords: On-policy distillation, Trajectory-Level KL Instability, Temporal Curriculum, multi-turn agent settings

Category: Reinforcement Learning

Research Objective:

– Investigate the instability and limitations of vanilla On-policy distillation (OPD) in multi-turn agent settings, particularly focusing on Trajectory-Level KL Instability.

Research Methods:

– Introduce TCOD (Temporal Curriculum On-Policy Distillation), employing a curriculum approach to progressively expand trajectory depth and enhance training stability.

Research Conclusions:

– TCOD effectively mitigates KL escalation, increases KL stability, and significantly improves agent performance in multi-turn tasks compared to vanilla OPD, with improvements up to 18 points. It can also surpass the teacher’s performance and generalize to tasks where the teacher fails.

Paper link: https://huggingface.co/papers/2604.24005

19. BARRED: Synthetic Training of Custom Policy Guardrails via Asymmetric Debate

Keywords: BARRED, custom guardrails, synthetic training data, multi-agent debate, dimension decomposition

Category: AI Systems and Tools

Research Objective:

– To develop a framework called BARRED for generating synthetic training data that enhances the performance of custom guardrail policies over existing language models.

Research Methods:

– Utilizes dimension decomposition and multi-agent debate to generate diverse and high-fidelity synthetic data without extensive human annotation.

Research Conclusions:

– The BARRED framework allows small language models to outperform state-of-the-art proprietary models by relying on synthetic data, highlighting the importance of dimension decomposition and debate-based verification for effective model fine-tuning.

Paper link: https://huggingface.co/papers/2604.25203

20. Step-Audio-R1.5 Technical Report

Keywords: Audio Language Models, Reinforcement Learning, Human Feedback, Chain-of-Thought, Immersive Dialogue

Category: Reinforcement Learning

Research Objective:

– Investigate the limitations of current reinforcement learning paradigms in audio language models, specifically addressing the “verifiable reward trap” and its impact on conversational quality.

Research Methods:

– Employ Reinforcement Learning from Human Feedback (RLHF) to refine audio reasoning capabilities, introducing Step-Audio-R1.5 to enhance interactive experiences.

Research Conclusions:

– Step-Audio-R1.5 effectively maintains analytical reasoning while transforming interaction quality, bridging the gap between mechanical verification and sensory empathy for immersive long-turn dialogues.

Paper link: https://huggingface.co/papers/2604.25719

21. Refinement via Regeneration: Enlarging Modification Space Boosts Image Refinement in Unified Multimodal Models

Keywords: Refinement via Regeneration, Unified multimodal models, text-to-image, semantic alignment, conditional image regeneration

Category: Multi-Modal Learning

Research Objective:

– To improve multi-modal model refinement by transitioning from editing-based approaches to conditional image regeneration, leading to better semantic alignment in text-to-image tasks.

Research Methods:

– A novel framework called Refinement via Regeneration is proposed that refines images by regenerating them based on conditional inputs, avoiding traditional editing methods.

Research Conclusions:

– Demonstrated significant improvement in evaluation metrics such as Geneval, DPGBench, and UniGenBench++, proving the efficacy of the RvR approach.

Paper link: https://huggingface.co/papers/2604.25636

22. AutoResearchBench: Benchmarking AI Agents on Complex Scientific Literature Discovery

Keywords: AI agents, AutoResearchBench, Deep Research, Wide Research, autonomous scientific research

Category: AI Systems and Tools

Research Objective:

– To present AutoResearchBench, a benchmark designed to evaluate AI agents’ capability in autonomous scientific literature discovery.

Research Methods:

– Utilizes two task types: Deep Research, involving multi-step probing, and Wide Research, which requires comprehensive paper collection.

Research Conclusions:

– AutoResearchBench sets a high difficulty benchmark, showing that powerful LLMs achieve low accuracy rates (9.39% in Deep Research and 9.31% in Wide Research) compared to previous benchmarks.

– Dataset, evaluation pipeline, and code are publicly released to encourage further research.

Paper link: https://huggingface.co/papers/2604.25256

23. Programming with Data: Test-Driven Data Engineering for Self-Improving LLMs from Raw Corpora

Keywords: Programming with Data, structured knowledge representation, language models, domain-specific capabilities, data repair

Category: Foundations of AI

Research Objective:

– To create a principled framework for systematically transferring human expertise into large language models using structured knowledge representation and systematic feedback.

Research Methods:

– Training data is treated as source code, enabling unit testing and debugging to address model failures identified as concept-level gaps and reasoning-chain breaks.

Research Conclusions:

– Demonstrates that the training data and model behavior relationship is traceable and repairable. This approach provides consistent improvements across different model scales and architectures without degrading general capabilities.

Paper link: https://huggingface.co/papers/2604.24819

1. Xiaomi open-sources MiMo-V2.5 series models under MIT license and launches Orbit 100 trillion token initiative

Xiaomi has officially open-sourced its MiMo-V2.5 series AI models under the permissive MIT license, allowing commercial use, inference deployment, and secondary training without additional authorization. The series includes MiMo-V2.5-Pro, optimized for Agent and coding tasks with 1 million token context window and ranked first among open-source models on GDPVal-AA and ClawEval benchmarks, and MiMo-V2.5, a native multimodal model supporting text, image, video, and audio understanding. Alongside the open-source release, Xiaomi launched the MiMo Orbit initiative, which includes a 100 trillion token giveaway program for AI builders over 30 days and an Agent ecosystem co-building program offering free token support for Agent framework teams. The models have achieved day-zero integration with multiple chip vendors including Alibaba Pingtouge, AWS Trainium, AMD, Baidu Kunlun, and inference frameworks SGLang and vLLM.

Read more: https://mp.weixin.qq.com/s/CX0mcpaE_sCTKxRBzebMvQ

Video Credit: The original article

2. QoderWork launches Expert Suite featuring 10 pre-built AI assistants for finance, legal, marketing and other professional domains

QoderWork has released Expert Suite, a new product feature that packages domain-specific AI capabilities into installable modules for enterprise users. The first batch includes 10 expert packages covering six areas: finance, legal, marketing, consulting, tax and accounting, and product management. Each suite combines multiple skills, tool connections, and standardized workflows into a single package that employees can install and use without technical configuration. The company also allows enterprises to create and distribute their own custom expert suites, enabling domain experts to encapsulate their workflows once for organization-wide deployment.

Read more: https://mp.weixin.qq.com/s/Z2T30ZbQXhzPIQxQxWJmxg

Video Credit: The original article

3. PixVerse integrates HappyHorse 1.0 video generation model with text-to-video and image-to-video capabilities

AISphere’s AI video generation platform PixVerse officially integrated the HappyHorse 1.0 video generation model on April 28. The model is available on PixVerse’s web platform, offering both text-to-video and image-to-video capabilities. HappyHorse 1.0 supports video generation up to 15 seconds in length, outputs at 1080P and 720P resolutions, and supports multiple aspect ratios including 16:9, 9:16, and 1:1. The integration expands PixVerse’s model ecosystem, which already includes its proprietary V, C, and R series models, providing users with more diverse AI video creation options on a unified platform.

Read more: https://mp.weixin.qq.com/s/Wu5Cx7K8dBqgufnQrAfB5w

Video Credit: @PixVerse_ on X

That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

Keywords: OneManCompany, multi-agent systems, agent identities, Talent Market, hierarchical loop

Category: AI Systems and Tools

Research Objective:

– The research aims to introduce OneManCompany (OMC), an organizational framework designed for multi-agent systems that facilitates dynamic team assembly, governance, and improvement through portable agent identities and hierarchical decision-making processes.

Research Methods:

– OMC encapsulates skills and runtime configurations into portable agent identities called Talents, coordinated through typed organizational interfaces. It utilizes an Explore-Execute-Review (E^2R) tree search for unified hierarchical decision-making, decomposing tasks into accountable units.

Research Conclusions:

– The OMC framework transforms static multi-agent systems into self-organizing and self-improving AI organizations, achieving significant improvements in success rate as demonstrated by its 84.67% success rate on PRDBench, surpassing the state of the art by 15.48 percentage points.

Paper link: https://huggingface.co/papers/2604.22446

2. ReVSI: Rebuilding Visual Spatial Intelligence Evaluation for Accurate Assessment of VLM 3D Reasoning

Keywords: ReVSI, Spatial Intelligence, Benchmark, VLMs, 3D Annotations

Category: Computer Vision

Research Objective:

– The paper aims to improve the validity of spatial intelligence evaluations by introducing a new benchmark, ReVSI, with enhanced annotations and controlled sampling conditions.

Research Methods:

– The research involved re-annotating objects and geometry across 381 scenes from five datasets, using professional 3D annotation tools, ensuring that each QA pair is correctly answerable under model inputs.

Research Conclusions:

– Evaluations using ReVSI reveal systematic failure modes in general and domain-specific VLMs that were hidden in previous benchmarks, leading to a more reliable assessment of spatial intelligence.

Paper link: https://huggingface.co/papers/2604.24300

3. Vision-Language-Action Safety: Threats, Challenges, Evaluations, and Mechanisms

Keywords: Vision-Language-Action models, embodied intelligence, adversarial attacks, data poisoning, safety challenges

Category: Robotics and Autonomous Systems

Research Objective:

– To provide a unified and comprehensive overview of safety in Vision-Language-Action models, addressing the unique safety challenges they present due to their embodied nature.

Research Methods:

– The survey organizes the safety aspects of VLA models along attack and defense timing axes, distinguishes VLA safety from other safety areas, and reviews threats, defenses, evaluations, and deployments in this domain.

Research Conclusions:

– The research highlights the fragmented nature of current literature and emphasizes the need for a unified approach to address safety challenges. It underlines key open problems such as establishing certified robustness for embodied trajectories and developing standardized evaluation and safety-aware training methods.

Paper link: https://huggingface.co/papers/2604.23775

4. Tuna-2: Pixel Embeddings Beat Vision Encoders for Multimodal Understanding and Generation

Keywords: Tuna-2, pixel embeddings, unified multimodal model, visual understanding, end-to-end optimization

Category: Multi-Modal Learning

Research Objective:

– Introduce Tuna-2, a unified multimodal model capable of performing visual understanding and generation directly from pixel embeddings, without relying on pretrained vision encoders.

Research Methods:

– Implementation of simple patch embedding layers to encode visual input, completely foregoing traditional modular vision encoder designs such as VAEs.

Research Conclusions:

– Tuna-2 achieves state-of-the-art performance in multimodal benchmarks.

– Demonstrates that end-to-end pixel-space learning offers scalable and stronger visual representations, particularly excelling in tasks requiring fine-grained visual perception.

– Highlights that pretrained vision encoders are unnecessary for effective multimodal modelling.

Paper link: https://huggingface.co/papers/2604.24763

5. World-R1: Reinforcing 3D Constraints for Text-to-Video Generation

Keywords: 3D constraints, reinforcement learning, video generation, geometric consistency, world simulation

Category: Generative Models

Research Objective:

– To enhance video generation by aligning it with 3D constraints using reinforcement learning and specialized text datasets.

Research Methods:

– Implemented World-R1 framework leveraging Flow-GRPO for structural coherence.

– Utilized feedback from pre-trained 3D foundation models and vision-language models.

– Adopted periodic decoupled training strategy to balance geometric consistency with scene fluidity.

Research Conclusions:

– The approach significantly improves 3D consistency while maintaining visual quality and scalability in video generation.

Paper link: https://huggingface.co/papers/2604.24764

1. GitHub Copilot transitions to usage-based billing model starting June 1, 2026

GitHub announced all Copilot plans will transition to usage-based billing on June 1, 2026, replacing premium request units with GitHub AI Credits based on token consumption. Base subscription prices remain unchanged, with monthly plans including AI Credits equivalent to their subscription cost. A preview billing experience launches in early May to help users understand projected costs before the transition takes effect.

Read more: https://github.blog/news-insights/company-news/github-copilot-is-moving-to-usage-based-billing/

Video Credit: The original article

2. OpenAI showcases interactive applications built with gpt-realtime-1.5 for voice-controlled app state management

OpenAI demonstrated interactive applications built with their gpt-realtime-1.5 model that enable users to control application state naturally through voice commands. The gpt-realtime-1.5 model is designed for low-latency speech-to-speech interactions and supports real-time voice agent capabilities. This showcases the practical implementation of voice-controlled interfaces that can manage application functionality through natural speech input.

Read more: https://github.com/openai/realtime-voice-component/

Video Credit: @OpenAIDevs on X

3. Microsoft launches Agent Mode for Copilot in Outlook with inbox and calendar management

Microsoft launched Agent Mode for Copilot in Outlook through the Frontier program beginning April 27. The new agentic capabilities enable Copilot to autonomously triage emails, reschedule meetings, manage calendar conflicts, and surface priority items before users ask. Agent Mode represents a shift from reactive assistance to proactive inbox and calendar management across Outlook for Windows and web.

Read more: https://techcommunity.microsoft.com/blog/outlook/copilot-in-outlook-new-agentic-experiences-for-email-and-calendar/4514601

Video Credit: @satyanadella on X

That’s all for today’s Global AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

Keywords: World Models, Predictive Environment Models, L1 Predictor, L2 Simulator, L3 Evolver

Category: Reinforcement Learning

Research Objective:

– Develop a taxonomy for world models to enhance predictive environment models for AI agents across multiple domains.

Research Methods:

– Introduce a “levels x laws” taxonomy with three capability levels (Predictor, Simulator, Evolver) and four law regimes (physical, digital, social, scientific).

– Synthesize over 400 works and summarize more than 100 representative systems in different AI applications.

Research Conclusions:

– The roadmap connects isolated research communities, moving from passive prediction to shaping environments through advanced world models.

– Proposal of decision-centric evaluation principles and a minimal reproducible evaluation package to aid in understanding and development.

Paper link: https://huggingface.co/papers/2604.22748

2. DiffNR: Diffusion-Enhanced Neural Representation Optimization for Sparse-View 3D Tomographic Reconstruction

Keywords: DiffNR, neural representation, CT reconstruction, single-step diffusion model, artifacts correction

Category: Computer Vision

Research Objective:

– The objective is to enhance neural representation optimization for CT reconstruction by integrating a novel framework called DiffNR, which addresses severe artifacts in sparse-view settings.

Research Methods:

– A single-step diffusion model, SliceFixer, is incorporated for artifact correction.

– Specialized conditioning layers are integrated, along with tailored data curation strategies for model fine-tuning.

– Pseudo-reference volumes are generated for auxiliary 3D perceptual supervision during reconstruction.

Research Conclusions:

– DiffNR significantly improves PSNR by 3.99 dB on average.

– It generalizes well across different domains and maintains efficient optimization, avoiding frequent diffusion model queries.

Paper link: https://huggingface.co/papers/2604.21518

3. Contexts are Never Long Enough: Structured Reasoning for Scalable Question Answering over Long Document Sets

Keywords: SLIDERS, Relational Database, SQL, Long Document Collections, Structured Reasoning

Category: Knowledge Representation and Reasoning

Research Objective:

– The objective is to enable scalable document question answering by extracting information into a relational database and using SQL-based structured reasoning.

Research Methods:

– The method involves extracting salient information into a relational database, applying SQL for reasoning, and introducing a data reconciliation stage for global coherence.

Research Conclusions:

– SLIDERS outperforms all baselines on three existing long-context benchmarks, exceeding strong base LLMs like GPT-4.1 by 6.6 points on average, and shows significant improvements on new benchmarks.

Paper link: https://huggingface.co/papers/2604.22294

4. AgentSearchBench: A Benchmark for AI Agent Search in the Wild

Keywords: AgentSearchBench, AI agents, execution-grounded signals, retrieval, reranking

Category: AI Systems and Tools

Research Objective:

– The study aims to address the challenge of identifying suitable AI agents for complex tasks by leveraging execution-grounded signals instead of relying solely on textual descriptions.

Research Methods:

– The research introduces AgentSearchBench, a benchmark evaluating agent search as retrieval and reranking problems. It utilizes nearly 10,000 real-world agents from multiple providers, assessing relevance through execution-grounded performance signals.

Research Conclusions:

– The study identifies a gap between semantic similarity and actual agent performance, highlighting limitations in description-based retrieval and reranking methods. It demonstrates that incorporating lightweight behavioral signals and execution-aware probing can significantly improve ranking quality in agent discovery.

Paper link: https://huggingface.co/papers/2604.22436

5. dWorldEval: Scalable Robotic Policy Evaluation via Discrete Diffusion World Model

Keywords: dWorldEval, Unified Token Space, Transformer-based Denoising, Robotics Policy Evaluation, Sparse Keyframe Memory

Category: Robotics and Autonomous Systems

Research Objective:

– To propose dWorldEval, a scalable evaluation proxy for robotics policies using a discrete diffusion world model.

Research Methods:

– Utilizes a discrete diffusion world model that maps vision, language, and robotic actions into a unified token space.

– Applies a transformer-based denoising network for modeling.

– Introduces sparse keyframe memory and progress token to maintain spatiotemporal consistency and evaluate task completion.

Research Conclusions:

– dWorldEval significantly outperforms previous approaches on various benchmarks, paving the way for advanced robotic evaluation.

Paper link: https://huggingface.co/papers/2604.22152

6. Memanto: Typed Semantic Memory with Information-Theoretic Retrieval for Long-Horizon Agents

Keywords: Agentic AI, Typed Semantic Memory, Information Theoretic Search, Knowledge Graph, Deterministic Retrieval

Category: AI Systems and Tools

Research Objective:

– To introduce Memanto, a universal memory layer for agentic AI that optimizes memory architecture by reducing computational overhead associated with traditional hybrid semantic graph architectures.

Research Methods:

– Implementation of Memanto using a typed semantic memory schema with predefined memory categories, an automated conflict resolution mechanism, and temporal versioning.

– Utilizing Moorcheh’s Information Theoretic Search engine for efficient data retrieval without indexing, benchmarking with LongMemEval and LoCoMo evaluation suites.

Research Conclusions:

– Memanto outperforms existing hybrid graph and vector-based systems in terms of accuracy and operational complexity, achieving high accuracy scores of 89.8% and 87.1% on evaluations, while maintaining lower complexity and requiring only a single retrieval query.

Paper link: https://huggingface.co/papers/2604.22085

7. DiagramBank: A Large-scale Dataset of Diagram Design Exemplars with Paper Metadata for Retrieval-Augmented Generation

Keywords: DiagramBank, AI scientist systems, schematic diagrams, multimodal retrieval, scientific figure generation

Category: Multi-Modal Learning

Research Objective:

– The primary goal of this research is to introduce DiagramBank, a large-scale dataset designed to bridge the gap in automated creation of publication-grade diagrams by AI scientist systems.

Research Methods:

– The researchers developed an automated curation pipeline that extracts figures and corresponding in-text references and employs a CLIP-based filter to differentiate schematic diagrams from standard plots or natural images.

Research Conclusions:

– DiagramBank, consisting of 89,422 schematic diagrams, is designed for enhanced multimodal retrieval and exemplar-driven scientific figure generation, enabling the effective synthesis of teaser figures for scientific manuscripts.

Paper link: https://huggingface.co/papers/2604.20857

8. Learning Evidence Highlighting for Frozen LLMs

Keywords: HiLight, Large Language Models, Reinforcement Learning, Emphasis Actor, Long-context reasoning

Category: Natural Language Processing

Research Objective:

– The study aims to enhance long-context reasoning in frozen large language models by introducing HiLight, which focuses on decoupling evidence selection from reasoning.

Research Methods:

– HiLight employs a lightweight Emphasis Actor trained through reinforcement learning, without the need for evidence labels or modifying the original solver, to insert highlight tags around key evidence.

Research Conclusions:

– HiLight consistently improves performance in tasks like sequential recommendation and long-context question answering across different solver sizes, demonstrating the Actor’s ability to capture genuine evidence structures.

Paper link: https://huggingface.co/papers/2604.22565

9.

Paper link: 

10. Emergent Strategic Reasoning Risks in AI: A Taxonomy-Driven Evaluation Framework

Keywords: Emergent Strategic Reasoning Risks, deception, reward hacking, ESRRSim, agentic framework

Category: Knowledge Representation and Reasoning

Research Objective:

– To systematically evaluate large language models for emergent strategic reasoning risks, including deception and reward hacking, using a taxonomy-driven framework.

Research Methods:

– Introduction of ESRRSim, an agentic framework, to assess reasoning traces and model responses across multiple LLMs, using a risk taxonomy with 7 categories and 20 subcategories.

Research Conclusions:

– The evaluation of 11 reasoning LLMs indicates substantial variation in risk profiles, with significant generational improvements, suggesting models may better recognize and adapt to evaluation contexts.

Paper link: https://huggingface.co/papers/2604.22119

11. AgriIR: A Scalable Framework for Domain-Specific Knowledge Retrieval

Keywords: Retrieval Augmented Generation, Modular Stages, AI for Agriculture, Language Models, Deterministic Citation

Category: AI Systems and Tools

Research Objective:

– Introduce AgriIR, a modular framework designed to access agricultural information efficiently through retrieval-augmented generation.

Research Methods:

– The framework utilizes modular stages such as query refinement, sub-query planning, retrieval, synthesis, and evaluation, ensuring adaptability to various knowledge verticals without changing architecture.

Research Conclusions:

– AgriIR demonstrates the capability to provide domain-accurate and trustworthy retrieval, even with limited resources, by emphasizing design and modular control. It exemplifies AI for Agriculture by promoting accessibility, sustainability, and accountability.

Paper link: https://huggingface.co/papers/2604.16353

12. EmbodiedMidtrain: Bridging the Gap between Vision-Language Models and Vision-Language-Action Models via Mid-training

Keywords: EmbodiedMidtrain, Vision-Language-Action Models, Mid-training, Robot Manipulation, Proximity Estimator

Category: Robotics and Autonomous Systems

Research Objective:

– Address the gap between Vision-Language Models and Vision-Language-Action Models to enhance robot manipulation performance through a mid-training approach.

Research Methods:

– Develop a mid-training data engine with a learnable proximity estimator to select VLA-aligned data from a VLM pool for improved downstream fine-tuning.

Research Conclusions:

– Mid-training effectively boosts performance across various VLM backbones, achieving competitive results with both expert VLAs and larger off-the-shelf VLMs. It provides a strong initialization for VLA fine-tuning, enhancing spatial reasoning while maintaining VLM data diversity.

Paper link: https://huggingface.co/papers/2604.20012

13. Sessa: Selective State Space Attention

Keywords: Sessa, attention, recurrent feedback path, power-law memory, selective retrieval

Category: Natural Language Processing

Research Objective:

– The study introduces Sessa, a decoder architecture designed to enhance long-context modeling by integrating attention within a recurrent feedback loop, offering an improvement over Transformers and state-space models.

Research Methods:

– Sessa integrates attention into a recurrent feedback path to create multiple attention-based paths, enhancing the influence of past tokens on future states with distinct power-law memory decay and flexible selective retrieval mechanisms.

Research Conclusions:

– Sessa outperforms other models in long-context benchmarks by achieving a power-law memory tail, ensuring superior performance and flexible retrieval compared to Transformer and Mamba-style baselines, while remaining competitive in short-context tasks.

Paper link: https://huggingface.co/papers/2604.18580

14. Building a Precise Video Language with Human-AI Oversight

Keywords: Video-language models, AI Native, Human-AI oversight, video captioning, video generation

Category: Multi-Modal Learning

Research Objective:

– The study aims to enhance Video-language models through structured visual specifications and a Human-AI oversight framework to improve captioning accuracy and enable better control over video generation.

Research Methods:

– They introduced open datasets, benchmarks, and a Critique-based Human-AI Oversight (CHAI) framework where experts critique and revise model-generated captions to ensure precision and recall in text generation. Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and inference-time scaling are employed to refine models.

Research Conclusions:

– The oversight framework significantly improves annotation accuracy, allowing open-source models to outperform closed-source counterparts. The methodology enables finer control over video generation, facilitating professional-level video understanding with applications in large-scale videos such as films and commercials. Data and code are accessible on their project page.

Paper link: https://huggingface.co/papers/2604.21718

15. FlowAnchor: Stabilizing the Editing Signal for Inversion-Free Video Editing

Keywords: FlowAnchor, video editing, Spatial-aware Attention Refinement, Adaptive Magnitude Modulation, video latent spaces

Category: Computer Vision

Research Objective:

– The main goal is to enable stable and efficient video editing by addressing signal instability in high-dimensional video latent spaces.

Research Methods:

– The approach introduces FlowAnchor, which utilizes Spatial-aware Attention Refinement to align textual guidance with spatial regions and Adaptive Magnitude Modulation to maintain editing strength.

Research Conclusions:

– FlowAnchor achieves temporally coherent and computationally efficient video editing, effectively handling challenging multi-object and fast-motion scenarios.

Paper link: https://huggingface.co/papers/2604.22586

16. LLM Safety From Within: Detecting Harmful Content with Internal Representations

Keywords: SIREN, Guard Models, Internal Features, Harmfulness Detection, Inference Efficiency

Category: AI Systems and Tools

Research Objective:

– To develop a lightweight guard model called SIREN that utilizes internal layer features from LLMs to enhance the detection efficiency and performance of harmful content.

Research Methods:

– Identification of safety neurons via linear probing

– Combination of neurons through an adaptive layer-weighted strategy

Research Conclusions:

– SIREN significantly outperforms state-of-the-art open-source guard models across multiple benchmarks with 250 times fewer trainable parameters.

– It exhibits superior generalization to unseen benchmarks and natural real-time streaming detection capabilities, improving inference efficiency compared to generative guard models.

Paper link: https://huggingface.co/papers/2604.18519

17. Video Analysis and Generation via a Semantic Progress Function

Keywords: Semantic Progress Function, Semantic Linearization, Semantic Pacing, Temporal Irregularities

Category: Generative Models

Research Objective:

– The primary aim is to develop a Semantic Progress Function to analyze and correct non-linear semantic evolution in media generated by models, improving transition smoothness through semantic linearization.

Research Methods:

– The Semantic Progress Function is introduced as a one-dimensional representation to capture semantic evolution in sequences. Semantic embeddings are used to compute distances for each frame, fitting a smooth curve to reflect cumulative semantic shifts.

Research Conclusions:

– The framework facilitates smoother transitions by reparameterizing sequences to ensure constant rate semantic change. It also serves as a model-agnostic tool to identify temporal irregularities and allows comparison of semantic pacing across different generators.

Paper link: https://huggingface.co/papers/2604.22554

1. GPT-5.5 by OpenAI

Product Hunt Data
Ranking: 12
Upvote: 396

Product Overview
GPT-5.5 is an AI-first model designed to run complex work end-to-end: it can plan, write and debug code, analyze data, synthesize research, and execute multi-step tasks by using tools and iterating with minimal direction, positioning the model as the primary operating layer for knowledge work.

Evaluation
AI Native Application Modernization: 92/100
Strong AI-native fit because the product’s core value is autonomous reasoning and orchestration rather than a UI feature; it modernizes workflows by replacing brittle scripts and manual coordination with a model that can decompose tasks, call tools, and refine outputs, with the main constraint being governance and reliability requirements for production use.

Website
https://openai.com/?ref=producthunt

2. DeepSeek-V4

Product Hunt Data
Ranking: 13
Upvote: 380

Product Overview
DeepSeek-V4 Preview is an AI-native MoE model family designed around long-context reasoning, offering V4-Pro and a lighter V4-Flash with a default 1M-token window. Its hybrid attention architecture makes ultra-long context practical by reducing compute and memory, enabling workflows like large-repo coding, multi-document analysis, and agent-style planning where the model is the core system.

Evaluation
AI Native Application Modernization: 89/100
DeepSeek-V4 modernizes AI applications by pushing the context layer to production scale, which can simplify retrieval-heavy pipelines and enable more stateful agents with fewer external components. The score reflects strong AI-native architecture and efficiency focus, while real-world modernization impact will still depend on model quality under 1M context, tooling integration, and operational controls for latency, cost, and safety.

Website
https://chat.deepseek.com/coder?ref=producthunt

3. Twenty 2.0

Product Hunt Data
Ranking: 15
Upvote: 367

Product Overview
Twenty 2.0 is an open-source CRM rebuilt as a programmable platform, letting teams define data models, objects, workflows, layouts, and widgets in code via an SDK that fits standard dev pipelines and AI-assisted tooling. AI is a core layer with support for custom agents and serverless functions, while remaining self-hostable so organizations can fully own deployment, customization, and governance.

Evaluation
AI Native Application Modernization: 87/100
The product modernizes CRM development by making configuration code-first and extensible, enabling AI agents to operate on first-class domain models and workflows rather than bolted-on automations. Strong self-hosting and customization support help enterprise control, with the main execution risk being the engineering lift to design robust models, permissions, and agent behaviors at scale.

Website
https://www.twenty.com/?ref=producthunt

4. Pegasus 1.5 by TwelveLabs

Product Hunt Data
Ranking: 34
Upvote: 203

Product Overview
Pegasus 1.5 is an API-first multimodal model that turns raw video into structured, timestamped metadata based on a domain-specific schema, making long-form footage (up to 2 hours) directly queryable and usable by downstream systems and agents, including reference-image based moments retrieval.

Evaluation
AI Native Application Modernization: 86/100
The product is AI-native because the model output (time-based structured metadata) becomes the core data layer for search, analytics, and automation workflows rather than a UI feature; strong fit for modernizing video operations into computable pipelines, with the main dependency being schema design quality and integration into existing data governance and retrieval systems.

Website
https://www.twelvelabs.io/?ref=producthunt

5. Lightfield

Product Hunt Data
Ranking: 45
Upvote: 672

Product Overview
Lightfield is an AI-native CRM that continuously builds and updates the system of record by reading emails, meetings, and calls, then lets teams query and act on that context in natural language to drive follow-ups, decks, and proposals without manual data entry.

Evaluation
AI Native Application Modernization: 86/100
The product treats AI as the primary data ingestion and reasoning layer—auto-creating CRM objects from real conversations and turning insights into drafted outputs—showing strong workflow automation and modernization, with remaining risk areas likely around data governance, accuracy on entity resolution, and safe integration into existing sales ops processes.

Website
https://lightfield.app/?ref=producthunt

6. Cai

Product Hunt Data
Ranking: 46
Upvote: 169

Product Overview
Cai is a local-first command layer that turns a single hotkey (⌥C) into an AI-native action runner across any on-screen context, letting you invoke prompts, scripts, and workflow actions like GitHub/Linear creation and route outputs to your preferred destinations without leaving your current app.

Evaluation
AI Native Application Modernization: 86/100
The product treats the model as the core execution engine rather than an add-on, with strong modernization signals via zero-setup local inference (bundled model), flexible model backends (MLX/HuggingFace and local/hosted connectors), and privacy-first defaults (no account/telemetry); integration breadth is solid, though deeper governance, team controls, and enterprise deployment patterns aren’t the primary focus.

Website
https://getcai.app/?ref=producthunt

7. Blink AI CFO

Product Hunt Data
Ranking: 53
Upvote: 163

Product Overview
Blink AI CFO is an AI-native finance operator inside Slack that can autonomously place stock and options trades via connected brokers and generate CFO-grade artifacts such as Excel financial models, live P&L sheets synced to Stripe and QuickBooks, and investor-ready slide decks on demand.

Evaluation
AI Native Application Modernization: 86/100
Core workflows are agent-driven end-to-end (trade execution, data syncing, model building, and deck generation) with real system integrations and artifact outputs, indicating true AI-native automation; remaining risk lies in governance, permissions, and auditability requirements for financial actions executed from chat.

Website
https://blink.new/?ref=producthunt

Statement: Evaluation results are generated by AI, lack of data support, reference learning only.

1. DeepSeek Releases Open-Source V4 Preview Models with 1M Context Length, Rivaling Top Closed-Source Models

On April 24, 2026, DeepSeek released preview versions of its V4 open-source AI models, including V4-Pro (1.6T total / 49B active parameters) and V4-Flash (284B total / 13B active parameters). Both models support a 1 million token context length and are available via the company’s API and chat interface. The models are designed to rival leading closed-source models in performance while remaining cost-effective and openly accessible to developers. DeepSeek has since reduced input cache pricing to 1/10th of the original price across its entire API series, with a 75% discount on V4-Pro running until May 5, 2026.

Read more: https://api-docs.deepseek.com/news/news260424

Video Credit: The original article

2. Kling AI Launches Native 4K Video Generation — Cinematic Quality in One Click

On April 24, 2026, Kling AI introduced a native 4K mode to its Video 3.0 series, enabling one-click 4K video generation with no super-resolution pipeline or complex post-processing required. Unlike tools that rely on upscaling, Kling’s 4K is generated natively, delivering cinematic-grade clarity and lighting in every frame. The feature is designed for large-screen display, high-definition playback, and professional production workflows, helping creators produce work with a true cinematic visual quality.

Read more: https://mp.weixin.qq.com/s/GHr4b7jeEzK71ZfLw1gdzA

Video Credit: @Kling_ai on X

3. Alibaba releases Qwen-Image-2.0-Pro text-to-image model with advanced typography and multilingual rendering

Alibaba announced the launch of Qwen-Image-2.0-Pro, an advanced text-to-image generation model that significantly improves image quality, multilingual text rendering, and instruction following. The model ranks 9th globally on AI Arena for text-to-image generation and offers enhanced consistency across different visual styles. Qwen-Image-2.0-Pro builds on the unified 7B parameter architecture that combines generation and editing capabilities, supporting native 2K resolution and professional typography rendering with up to 1,000-token prompts.

Read more: https://x.com/i/web/status/2048022731548229869

Video Credit: @Alibaba_Qwen on X

That’s all for today’s China AI Native Industry Insights. Join us at AI Native Foundation Membership Dashboard for the latest insights on AI Native, or follow our linkedin account at AI Native Foundation and our twitter account at AINativeF.

Keywords: Time Series Reasoning, TSRM, Chain-of-Thought, Vision-Language Models, Multi-stage Curriculum Fine-tuning

Category: Knowledge Representation and Reasoning

Research Objective:

– The study aims to improve the understanding of temporal data by Large Language Models (LLMs) through the introduction of a hierarchical time series reasoning dataset and model.

Research Methods:

– The researchers formulated a four-level taxonomy to enhance cognitive complexity in Time Series Reasoning and introduced HiTSR, a dataset with 83k samples.

– Proposed the TSRM LLaTiSA employing visualized patterns and precision-calibrated numerical tables, fine-tuned via a multi-stage curriculum approach.

Research Conclusions:

– LLaTiSA demonstrates superior performance and strong out-of-distribution generalization across various TSR tasks and real-world scenarios.

Paper link: https://huggingface.co/papers/2604.17295

2. UniT: Toward a Unified Physical Language for Human-to-Humanoid Policy Learning and World Modeling

Keywords: UniT, Human-to-Humanoid Transfer, Cross-Reconstruction Mechanism, Embodiment-Agnostic, Shared Latent Space

Category: Robotics and Autonomous Systems

Research Objective:

– To create a unified visual-language representation that enables efficient and scalable human-to-humanoid action transfer by overcoming kinematic mismatches.

Research Methods:

– The introduction of UniT framework utilizing a tri-branch cross-reconstruction mechanism to anchor kinematics to physical outcomes and a fusion branch for a shared latent space of embodiment-agnostic intents.

Research Conclusions:

– UniT offers significant improvements in data efficiency and generalization in policy learning, demonstrating zero-shot task transfer in simulations and real-world. It also enables direct cross-embodiment action transfer, enhancing humanoid video generation through aligned dynamics.

Paper link: https://huggingface.co/papers/2604.19734

3. Co-Evolving LLM Decision and Skill Bank Agents for Long-Horizon Tasks

Keywords: Large Language Models, Skill Bank, Skill Pipeline, Long Horizon Decision Making

Category: Reinforcement Learning

Research Objective:

– To develop a co-evolution framework called COSPLAY that enhances Large Language Models’ ability to discover, retain, and reuse structured skills across episodes in long-horizon interactive environments.

Research Methods:

– Implementation of a learnable skill bank and a skill pipeline to guide LLM decision agents, enabling improved skill retrieval and action generation.

Research Conclusions:

– COSPLAY framework, using an 8B base model, achieves over 25.1% average reward improvement against four frontier LLM baselines on single-player game benchmarks and remains competitive on multi-player social reasoning games.

Paper link: https://huggingface.co/papers/2604.20987

4. VLAA-GUI: Knowing When to Stop, Recover, and Search, A Modular Framework for GUI Automation

Keywords: GUI agentic framework, Completeness Verifier, Loop Breaker, Search Agent, Coding Agent

Category: AI Systems and Tools

Research Objective:

– The objective is to tackle early stopping and repetitive loop issues in autonomous GUI agents using a modular framework called VLAA-GUI, which integrates components for verification, loop breaking, and search capabilities.

Research Methods:

– The study employs three integrated components: Completeness Verifier for enforcing success criteria, Loop Breaker for managing interaction modes and strategy shifts, and Search Agent for querying LLMs for unfamiliar workflows. These components are further supported by Coding and Grounding Agents when needed.

Research Conclusions:

– Evaluation across five backbones on Linux and Windows tasks showed top performance, with some models surpassing human performance. Ablation studies reveal consistent improvement across frameworks, particularly benefiting weaker backbones with sufficient step budgets. The Loop Breaker significantly reduces step wastage for loop-prone models.

Paper link: https://huggingface.co/papers/2604.21375

5. Hybrid Policy Distillation for LLMs

Keywords: Knowledge Distillation, Large Language Models, Hybrid Policy Distillation, Mode Coverage, Computational Efficiency

Category: Natural Language Processing

Research Objective:

– To improve the stability and efficiency of knowledge distillation across various model sizes and tasks using a novel approach called Hybrid Policy Distillation.

Research Methods:

– Combines forward and reverse KL divergence to balance mode coverage and mode-seeking.

– Utilizes off-policy data with lightweight, approximate on-policy sampling.

Research Conclusions:

– Demonstrates improved optimization stability and computational efficiency.

– Shows enhanced performance in long-generation math reasoning, short-generation dialogue, and code tasks across diverse model families and scales.

Paper link: https://huggingface.co/papers/2604.20244

6. EditCrafter: Tuning-free High-Resolution Image Editing via Pretrained Diffusion Model

Keywords: EditCrafter, high-resolution image editing, pretrained text-to-image diffusion models, tiled inversion, noise-damped manifold-constrained guidance

Category: Generative Models

Research Objective:

– The objective of the research is to introduce EditCrafter, a method for editing high-resolution images without the need for tuning, by utilizing pretrained text-to-image diffusion models.

Research Methods:

– The approach involves the use of tiled inversion to maintain the original identity of high-resolution images and a proposed noise-damped manifold-constrained classifier-free guidance (NDCFG++) to achieve effective image editing.

Research Conclusions:

– EditCrafter has been shown to produce impressive editing results across various resolutions without the need for fine-tuning or optimization.

Paper link: https://huggingface.co/papers/2604.10268

7. Vista4D: Video Reshooting with 4D Point Clouds

Keywords: Vista4D, 4D point cloud, video reshooting, camera control, 4D consistency

Category: Computer Vision

Research Objective:

– Introduce Vista4D, a video reshooting framework leveraging 4D point cloud representation to synthesize scenes from different viewpoints while ensuring 4D consistency and camera control.

Research Methods:

– Utilize a 4D-grounded point cloud representation with static pixel segmentation and 4D reconstruction to maintain content appearance and provide rich camera signals.

– Train with reconstructed multiview dynamic data for robust performance against point cloud artifacts during real-world inference.

Research Conclusions:

– Demonstrated improved 4D consistency, camera control, and visual quality compared to state-of-the-art baselines across various videos and camera paths.

– Method generalizes to applications like dynamic scene expansion and 4D scene recomposition.

Paper link: https://huggingface.co/papers/2604.21915

8. Encoder-Free Human Motion Understanding via Structured Motion Descriptions

Keywords: Structured Motion Description, text-based large language models, motion question answering, motion captioning, LoRA adaptation

Category: Natural Language Processing

Research Objective:

– Introduce Structured Motion Description (SMD) to enhance large language models (LLMs) in human motion reasoning by converting joint position sequences into structured natural language.

Research Methods:

– Employ a rule-based, deterministic approach inspired by biomechanical analysis, transforming motion data into descriptive text without requiring learned encoders or alignment modules.

Research Conclusions:

– SMD surpasses state-of-the-art results in motion question answering and captioning tasks and offers practical benefits like interoperability across different LLMs and human-readable representation for interpretable attention analysis.

Paper link: https://huggingface.co/papers/2604.21668

9. Trust but Verify: Introducing DAVinCI — A Framework for Dual Attribution and Verification in Claim Inference for Language Models

Keywords: Large Language Models, Dual Attribution, Verification framework, factual reliability, entailment-based reasoning

Category: Natural Language Processing

Research Objective:

– To enhance the factual reliability and interpretability of Large Language Model outputs through a dual attribution and verification framework called DAVinCI.

Research Methods:

– DAVinCI operates in two stages: it attributes claims to internal components and external sources, and it verifies these claims using entailment-based reasoning and confidence calibration.

Research Conclusions:

– DAVinCI significantly improves classification accuracy, attribution precision, recall, and F1-score by 5-20% compared to standard verification-only baselines. It provides a scalable solution for building auditable and trustworthy AI systems.

Paper link: https://huggingface.co/papers/2604.21193

10. PersonalAI: A Systematic Comparison of Knowledge Graph Storage and Retrieval Approaches for Personalized LLM agents

Keywords: knowledge graph, external memory framework, Retrieval-Augmented Generation, large language models, temporal dependencies

Category: Natural Language Processing

Research Objective:

– To enhance language model personalization by effectively incorporating user interaction history through a knowledge graph-based external memory framework.

Research Methods:

– Implementation of a flexible external memory framework using a knowledge graph constructed automatically by large language models.

– Introduction of a hybrid graph design supporting dynamic semantic and temporal representations with varied retrieval mechanisms like A*, WaterCircles traversal, and beam search.

Research Conclusions:

– Demonstrated that different memory and retrieval configurations yield optimal performance across various tasks such as TriviaQA and DiaASQ benchmarks.

– Extended DiaASQ with temporal annotations and internally contradictory statements to validate the system’s robustness in managing temporal dependencies and context-aware reasoning.

Paper link: https://huggingface.co/papers/2506.17001

11. Temporally Extended Mixture-of-Experts Models

Keywords: Mixture-of-Experts, Reinforcement Learning, Options Framework, Deliberation Costs

Category: Reinforcement Learning

Research Objective:

– To reduce expert switching rates in Mixture-of-Experts models while maintaining model accuracy using a temporal extension via the reinforcement learning options framework.

Research Methods:

– Utilization of the options framework in reinforcement learning, integrating a controller in the model to determine when and which expert sets to switch, applying the method to models like gpt-oss-20b.

Research Conclusions:

– The approach significantly reduces expert switch rates from over 50% to below 5% while retaining up to 90% of base-model accuracy, illustrating the feasibility of converting pre-trained models to temporally extended MoEs with minimal training.

Paper link: https://huggingface.co/papers/2604.20156

12.

Paper link: 

13. 3D-VCD: Hallucination Mitigation in 3D-LLM Embodied Agents through Visual Contrastive Decoding

Keywords: 3D-VCD, Hallucination Mitigation, Visual Contrastive Decoding, Geometric Perturbations, 3D Scene Graph

Category: Knowledge Representation and Reasoning

Research Objective:

– The main objective is to introduce 3D-VCD, the first inference-time visual contrastive decoding framework focused on mitigating hallucinations in 3D embodied agents.

Research Methods:

– 3D-VCD constructs distorted 3D scene graphs through semantic and geometric perturbations applied to object-centric representations, contrasting predictions between original and perturbed contexts.

Research Conclusions:

– The study demonstrates that 3D-VCD effectively improves grounded reasoning without retraining, suggesting it as a practical solution for enhancing the reliability of embodied intelligence by using inference-time contrastive decoding with structured 3D representations.

Paper link: https://huggingface.co/papers/2604.08645

14. Test-Time Adaptation for EEG Foundation Models: A Systematic Study under Real-World Distribution Shifts

Keywords: EEG foundation models, AI-generated summary, Test-time adaptation, Optimization-free methods, Gradient-based approaches

Category: AI in Healthcare

Research Objective:

– To introduce NeuroAdapt-Bench for systematically evaluating test-time adaptation methods on EEG foundation models under realistic distribution shifts.

Research Methods:

– The study evaluates representative TTA approaches from other domains across multiple pretrained foundation models, diverse downstream tasks, and heterogeneous datasets that include in-distribution, out-of-distribution, and extreme modality shifts like Ear-EEG.

Research Conclusions:

– Standard TTA methods show inconsistent and often degrading performance, with gradient-based approaches particularly affected, whereas optimization-free methods offer greater stability and more reliable improvements.

– The findings highlight the limitations of existing TTA techniques in EEG and emphasize the need for domain-specific adaptation strategies.

Paper link: https://huggingface.co/papers/2604.16926

15. Explainable Disentangled Representation Learning for Generalizable Authorship Attribution in the Era of Generative AI

Keywords: Variational Autoencoder, Authorship Attribution, AI-generated Text Detection, Discriminative Disentanglement, Explainable AI

Category: Natural Language Processing

Research Objective:

– To improve authorship attribution and AI-generated text detection by disentangling style from content using a novel framework called Explainable Authorship Variational Autoencoder (EAVAE).

Research Methods:

– Utilized supervised contrastive learning for pretraining style encoders.

– Implemented a variational autoencoder architecture with separate encoders for style and content representations.

– Employed a novel discriminator for effective disentanglement and generation of natural language explanations.

Research Conclusions:

– Presented method achieves state-of-the-art performance in authorship attribution across various datasets.

– Demonstrated superior performance in AI-generated text detection, especially in few-shot learning scenarios.

Paper link: https://huggingface.co/papers/2604.21300

16. Coevolving Representations in Joint Image-Feature Diffusion

Keywords: CoReDi, semantic representation, VAE latents, generative modeling, lightweight linear projection

Category: Generative Models

Research Objective:

– The study aims to enhance generative modeling by adapting the semantic representation space during training through Coevolving Representation Diffusion (CoReDi).

Research Methods:

– CoReDi utilizes a lightweight linear projection that evolves alongside the diffusion model and employs techniques such as stop-gradient targets, normalization, and targeted regularization to stabilize the process.

Research Conclusions:

– CoReDi improves convergence speed and sample quality in generative models using both VAE latent and pixel-space diffusion, demonstrating better performance than models with fixed representation spaces.

Paper link: https://huggingface.co/papers/2604.17492

17. WebGen-R1: Incentivizing Large Language Models to Generate Functional and Aesthetic Websites with Reinforcement Learning

Keywords: reinforcement learning, website generation, multi-page websites, multimodal rewards, Large Language Models

Category: Reinforcement Learning

Research Objective:

– The study aims to develop a novel framework, WebGen-R1, for project-level website generation using reinforcement learning that integrates structured scaffolding and multimodal rewards.

Research Methods:

– Utilizes an end-to-end RL framework with a scaffold-driven generation paradigm and introduces cascaded multimodal rewards to enhance architectural integrity and aesthetics in multi-page website creation.

Research Conclusions:

– WebGen-R1 significantly outperforms existing open-source models in generating deployable, visually aesthetic websites and rivals state-of-the-art models in functional success and rendering validity.

Paper link: https://huggingface.co/papers/2604.20398

18. UniGenDet: A Unified Generative-Discriminative Framework for Co-Evolutionary Image Generation and Generated Image Detection

Keywords: AI-generated summary, generative networks, discriminative frameworks, generative-discriminative framework, multimodal self-attention mechanism

Category: Generative Models

Research Objective:

– The primary goal of the research is to develop a unified generative-discriminative framework, called UniGenDet, that enables co-evolutionary image generation and detection using symbiotic attention mechanisms and unified fine-tuning algorithms.

Research Methods:

– The study proposes a symbiotic multimodal self-attention mechanism and a unified fine-tuning algorithm to bridge the task gap between image generation and detection.

– It introduces a detector-informed generative alignment mechanism to enhance seamless information exchange.

Research Conclusions:

– The proposed UniGenDet framework allows for improved interpretability in authenticity identification and guides the creation of higher-fidelity images.

– Extensive experiments demonstrate that the method achieves state-of-the-art performance across multiple datasets.

Paper link: https://huggingface.co/papers/2604.21904

19. Context Unrolling in Omni Models

Keywords: Omni, multimodal model, Context Unrolling, multimodal knowledge manifold, downstream reasoning

Category: Multi-Modal Learning

Research Objective:

– Introduce Omni, a unified model designed to improve reasoning capabilities across various data modalities.

Research Methods:

– Omni is trained natively on diverse modalities such as text, images, videos, 3D geometry, and hidden representations. This enables the model to perform Context Unrolling, allowing it to reason explicitly across multiple modal representations before predictions.

Research Conclusions:

– Omni enhances the aggregation of complementary information from heterogeneous modalities, leading to a more accurate approximation of the shared multimodal knowledge manifold. The model delivers strong performance on generation and understanding benchmarks and showcases advanced reasoning capabilities across modalities.

Paper link: https://huggingface.co/papers/2604.21921

20. TingIS: Real-time Risk Event Discovery from Noisy Customer Incidents at Enterprise Scale

Keywords: incident discovery, event linking engine, Large Language Models, cascaded routing, noise reduction

Category: AI Systems and Tools

Research Objective:

– The primary aim is to develop TingIS, an enterprise-grade system that efficiently identifies critical incidents from high-volume, noisy customer reports in cloud-native services.

Research Methods:

– The system employs a multi-stage event linking engine using Large Language Models to merge events effectively. It also uses a cascaded routing mechanism for precise business attribution and a multi-dimensional noise reduction pipeline.

Research Conclusions:

– TingIS, when deployed in a real-world environment, achieves a P90 alert latency of 3.5 minutes and a 95% discovery rate for high-priority incidents, significantly surpassing baseline methods in routing accuracy, clustering quality, and Signal-to-Noise Ratio.

Paper link: https://huggingface.co/papers/2604.21889

21. Seeing Fast and Slow: Learning the Flow of Time in Videos

Keywords: Video Speed Manipulation, Self-Supervised Learning, Temporal Reasoning, Slow-Motion Video, High-Speed Cameras

Category: Computer Vision

Research Objective:

– To explore and manipulate the flow of time in videos as a learnable visual concept.

Research Methods:

– Developed self-supervised temporal reasoning models to detect speed changes and estimate video playback speed using multimodal cues and temporal structure.

– Curated the largest slow-motion video dataset from in-the-wild sources for further model development.

Research Conclusions:

– Demonstrated temporal control capabilities, including speed-conditioned video generation and temporal super-resolution, transforming low-FPS videos into high-FPS sequences.

– Highlighted the potential for temporally controllable video generation, temporal forensics detection, and enriched world-models understanding event unfolding over time.

Paper link: https://huggingface.co/papers/2604.21931

22. StyleID: A Perception-Aware Dataset and Metric for Stylization-Agnostic Facial Identity Recognition

Keywords: facial identity preservation, stylization, human perception-aware dataset, semantic encoders

Category: Computer Vision

Research Objective:

– The primary aim of the study is to evaluate and preserve facial identity under various creative stylizations, such as cartoons and paintings, with an emphasis on aligning algorithms with human perception.

Research Methods:

– The research employed StyleID, a comprehensive evaluation framework comprising two datasets, StyleBench-H and StyleBench-S, derived from both human verification judgments and psychometric experiments to ensure the preservation of facial identity across different styles and strengths.

Research Conclusions:

– Findings indicate that the use of calibrated semantic encoders in the StyleID framework enhances the correlation with human judgments and robustness in recognizing facial identities in artist-drawn portraits, demonstrating the need for a style-agnostic framework in facial identity preservation.

Paper link: https://huggingface.co/papers/2604.21689

23. WorldMark: A Unified Benchmark Suite for Interactive Video World Models

Keywords: Interactive Video Generation, WorldMark Benchmark, Unified Controls, Evaluation Metrics, World Models

Category: Generative Models

Research Objective:

– Introduce WorldMark, a standardized benchmark for evaluating interactive video generation models under identical scenarios and controls.

Research Methods:

– Develop a unified action-mapping layer for consistent model comparisons.

– Create a hierarchical test suite with varied evaluation cases.

– Offer a modular evaluation toolkit for diverse visual quality metrics.

Research Conclusions:

– WorldMark enables fair cross-model comparisons by standardizing test conditions.

– Launches World Model Arena for real-time model competition and evaluation.

Paper link: https://huggingface.co/papers/2604.21686