🔥 Ideas Have Genomes: Benchmarking Scientific Lineage Reasoning and Lineage-Grounded Idea Generation

💡 The paper introduces a benchmark for evaluating the ability of artificial intelligence systems to reason about and generate scientific ideas based on their lineage. The authors propose that scientific ideas can be represented as genetic-like objects, called Idea Genome objects, which capture the inheritance structure of ideas. The benchmark, called IdeaGene-Bench, consists of a large dataset of scientific papers and proposals, represented as Idea Genome objects, and includes records of how these objects evolve over time.

The benchmark supports two types of evaluations: IG-Exam, which tests a system's ability to reason about the lineage of ideas, and IG-Arena, which evaluates a system's ability to generate new ideas that are coherent with a given lineage. The IG-Exam evaluation includes 42 task types and 1029 instances, and tests a system's ability to perform tasks such as tracing the inheritance of ideas, reasoning about evolutionary relationships, and verifying the lineage of a proposal.

The IG-Arena evaluation uses a lineage-conditioned Population-Evolution Score to assess a system's ability to generate new ideas that are meaningful and valuable. The score takes into account whether a proposal inherits the right Idea Genome objects, varies meaningfully from nearby work, and offers selection value for future research.

The authors conducted experiments on 14 language model-based systems and found that the strongest system achieved only 27.3% exact accuracy on lineage reasoning tasks. The results also showed that providing structured lineage context did not uniformly improve the performance of all systems, but rather reshuffled the rankings of the systems.

The paper's contributions include the introduction of the IdeaGene-Bench benchmark, the Idea Genome framework for representing scientific ideas, and the evaluation of language model-based systems on lineage reasoning and idea generation tasks. The results highlight the challenges of developing systems that can reason about and generate scientific ideas based on their lineage, and demonstrate the need for further research in this area.

📅 Published on Jul 9

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2607.08758
• PDF: https://arxiv.org/pdf/2607.08758
• Project Page: https://visionxlab.github.io/IdeasHaveGenomes/

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

#IdeaGeneration #ScientificLineage #ArtificialIntelligenceInScience #IdeaGenome #LineageReasoning

🔥 Advancing Open-source World Models

💡 The paper presents LingBot-World, an open-source world simulator that offers high-fidelity dynamics, long-term memory capabilities, and real-time interactivity across diverse environments. The simulator is capable of generating realistic and scientifically accurate environments, as well as cartoon styles and other settings. It achieves a high level of fidelity and robustness in its dynamics, allowing it to maintain contextual consistency over time, which is referred to as long-term memory. This means that the simulator can preserve the context of a scene over a period of minutes, rather than just seconds. Additionally, LingBot-World supports real-time interactivity, with a latency of under 1 second when producing 16 frames per second. The release of LingBot-World as an open-source model aims to bridge the gap between open-source and closed-source technologies, and is expected to have practical applications in areas such as content creation, gaming, and robot learning. The authors provide public access to the code and model, allowing the community to build upon and extend the capabilities of LingBot-World. Overall, the paper contributes to the advancement of open-source world models, providing a powerful tool for researchers and developers to explore and create complex simulated environments.

📅 Published on Jan 28

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2601.20540
• PDF: https://arxiv.org/pdf/2601.20540
• Project Page: https://technology.robbyant.com/lingbot-world

🤖 Models citing this paper:
• https://huggingface.co/robbyant/lingbot-world-base-cam
• https://huggingface.co/robbyant/lingbot-world-base-act-preview
• https://huggingface.co/robbyant/lingbot-world-fast

🚀 Spaces citing this paper:
• https://huggingface.co/spaces/naeem123goplay/Real-World
• https://huggingface.co/spaces/iCoderX/Lingbot-World
• https://huggingface.co/spaces/ftdgfxxfg66/Real-World

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

#OpenSourceWorldModels #WorldSimulators #LongTermMemoryInAI #RealTimeInteractivity #HighFidelityDynamics

🔥 Vidu S1: A Real-Time Interactive Video Generation Model

💡 The paper introduces Vidu S1, a real-time interactive video generation model that enables voice-controlled digital character animation with high frame rates on consumer hardware. The model addresses the problem of generating high-quality, real-time video content that can be controlled by users through voice instructions. To achieve this, the authors employ TurboDiffusion and TurboServe, which allow Vidu S1 to produce 540p videos at up to 42 frames per second on regular consumer GPUs. The model supports infinite-length video generation without visual distortion and allows users to upload custom images and choose different voice tones for personalized experiences. The results show that Vidu S1 achieves the best performance across all test metrics while meeting real-time inference requirements, demonstrating its effectiveness in generating high-quality, interactive video content. Overall, the paper presents a significant contribution to the field of video generation, enabling real-time, voice-controlled, and personalized video experiences.

📅 Published on Jul 3

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2607.03118
• PDF: https://arxiv.org/pdf/2607.03118
• Project Page: https://vidu.com/vidu-stream

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

#RealTimeVideoGeneration #InteractiveVideoModel #VoiceControlledAnimation #DigitalCharacterAnimation #TurboDiffusionTechnology

🔥 CommonForms: A Large, Diverse Dataset for Form Field Detection

💡 This paper introduces CommonForms, a large and diverse dataset for form field detection, which involves identifying the location and type of form fields such as text input, choice buttons, and signatures in images of pages. The dataset was constructed by filtering 8 million documents from Common Crawl to arrive at a final dataset of 55,000 documents with over 450,000 pages, containing a diverse mixture of languages and domains. The dataset is unique in that one third of the pages are non-English and no single domain makes up more than 25 percent of the dataset.

The paper also presents a family of form field detectors, FFDNet-Small and FFDNet-Large, which achieve high average precision on the CommonForms test set. The models were trained at a relatively low cost of less than 500 dollars. The results show that high-resolution inputs are crucial for high-quality form field detection and that the cleaning process improves data efficiency.

The paper makes several key contributions, including the release of the first large-scale dataset for form field detection and the first open-source models. The dataset, models, and code will be made available to the public. The qualitative analysis shows that the models outperform a popular commercially available PDF reader and can predict checkboxes in addition to text and signature fields, making this a significant advancement in the field of form field detection. Overall, the paper provides a valuable resource for researchers and developers working on form field detection and related tasks.

📅 Published on Sep 20, 2025

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2509.16506
• PDF: https://arxiv.org/pdf/2509.16506

🤖 Models citing this paper:
• https://huggingface.co/jbarrow/FFDNet-L
• https://huggingface.co/jbarrow/FFDNet-L-cpu
• https://huggingface.co/jbarrow/FFDNet-S

📊 Datasets citing this paper:
• https://huggingface.co/datasets/Voxel51/commonforms_val_subset
• https://huggingface.co/datasets/jbarrow/CommonForms
• https://huggingface.co/datasets/WEwoCram/CommonForms

🚀 Spaces citing this paper:
• https://huggingface.co/spaces/jbarrow/CommonForms
• https://huggingface.co/spaces/Voxel51/2025-ai-timeline
• https://huggingface.co/spaces/PSynx/widget-detector-demo

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

#FormFieldDetection #DocumentImageAnalysis #FormFieldRecognition #DatasetConstruction #DocumentUnderstanding

🔥 Hunyuan3D 2.0: Scaling Diffusion Models for High Resolution Textured 3D Assets Generation

💡 The paper introduces Hunyuan3D 2.0, a large-scale 3D synthesis system for generating high-resolution textured 3D assets. The system consists of two main components: Hunyuan3D-DiT, a shape generation model, and Hunyuan3D-Paint, a texture synthesis model. The shape generation model uses a scalable flow-based diffusion transformer to create geometry that aligns with a given condition image. The texture synthesis model uses strong geometric and diffusion priors to produce high-resolution and vibrant texture maps for generated or hand-crafted meshes.

The authors also introduce Hunyuan3D-Studio, a user-friendly production platform that simplifies the re-creation process of 3D assets, allowing both professional and amateur users to manipulate or animate their meshes efficiently. The system is evaluated and compared to previous state-of-the-art models, showing that Hunyuan3D 2.0 outperforms them in terms of geometry details, condition alignment, and texture quality.

The main contributions of the paper are the development of a scalable and efficient 3D synthesis system, the introduction of a user-friendly production platform, and the public release of the code and pre-trained weights of the models. The system aims to fill the gaps in the open-source 3D community for large-scale foundation generative models, providing a valuable resource for researchers and developers. Overall, the paper presents a significant advancement in the field of 3D synthesis, enabling the generation of high-quality textured 3D assets with improved geometry and texture details.

📅 Published on Jan 21, 2025

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2501.12202
• PDF: https://arxiv.org/pdf/2501.12202
• Project Page: https://huggingface.co/alluriaiprojects

🤖 Models citing this paper:
• https://huggingface.co/tencent/Hunyuan3D-2
• https://huggingface.co/tencent/Hunyuan3D-2.1
• https://huggingface.co/tencent/Hunyuan3D-2mv

📊 Datasets citing this paper:
• https://huggingface.co/datasets/tencent/HY3D-Bench

🚀 Spaces citing this paper:
• https://huggingface.co/spaces/tencent/Hunyuan3D-2
• https://huggingface.co/spaces/frogleo/Image-to-3D
• https://huggingface.co/spaces/HorizonRobotics/EmbodiedGen-Image-to-3D

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

#DiffusionModels #3DAssetGeneration #Textured3DModeling #GeometrySynthesis #3DSynthesisSystems

🔥 OPSD-V: On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators

💡 The paper proposes a method called On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators, or OPSD-V, which aims to improve the quality of videos generated by few-step autoregressive video diffusion models. The problem with existing models is that they can produce long videos with low latency, but the quality of the video degrades over time due to error accumulation and weakened motion dynamics.

To address this issue, OPSD-V introduces real long-video data as temporal context during training, providing dense trajectory-level supervision to improve visual quality and motion dynamics. The method works by having a student model follow the exact inference-time rollout, generating each chunk of the video conditioned on its own previously generated cache. In parallel, a teacher model is evaluated at the same denoising states, but uses a cleaner temporal cache that can be replaced by real-video context. This provides corrective targets under on-policy cache dynamics, without changing the inference mechanism.

The results show that OPSD-V consistently improves the visual quality, motion dynamics, and VBenchLong scores of the generated videos. The method is applied to representative few-step autoregressive video models, including Self-Forcing and LongLive, and the experiments demonstrate significant improvements. A user study with 10 participants also shows that OPSD-V is preferred over the base models in 66 percent of overall-preference judgments, and 82.5 percent excluding ties. Overall, the paper contributes a novel method for improving the quality of videos generated by few-step autoregressive video diffusion models, without altering the inference mechanism.

📅 Published on Jul 9

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2607.08766
• PDF: https://arxiv.org/pdf/2607.08766
• Project Page: https://meigen-ai.github.io/OPSD-V/

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

#AutoregressiveVideoGeneration #VideoDiffusionModels #SelfDistillationTechniques #FewStepVideoGeneration #PostTrainingOptimization

Create your own AI assistant for free in 5 minutes.

It's a familiar problem: everyone wants a personal AI assistant, but building one from scratch usually means servers, API keys, integrations, maintenance, and a ton of technical overhead.

Amplify takes care of all of this for you. In about 5 minutes, you'll have a personal AI agent connected to your Google account—Gmail, Drive, Calendar, Docs, Slides, Sheets, and more. Google integration is officially verified.

🗣You can communicate with your assistant anywhere: Telegram, WhatsApp, Slack, WeChat, or Discord.

It can help with email, draft replies to text or voice messages, send emails, set reminders, create and manage spreadsheets, generate images, create videos, edit short videos, work with PDFs, Notion, Obsidian, and much more.

Dozens of skills are already available, and the list is constantly growing. If you need a custom skill for your workflow, business, or team, the Amplify team will quickly develop and implement it.

The pricing is simple: $10 per month plus pay only for the features you actually use. No confusing token system—the cost of each action is clearly displayed in your dashboard.

And if you already have a ChatGPT subscription, you can sign up and essentially avoid paying separately for the AI ​​model.

😎For subscribers: use the promo code and get two months free + $10 credit to your balance.

After registering, you'll receive your own promo code. If someone else signs up with it, you'll get an extra month free.

Try Amplify here: https://getamplify.team/
Promo code: CODEPROGRAMMER

🔥 Long-Horizon-Terminal-Bench: Testing the Limits of Agents on Long-Horizon Terminal Tasks with Dense Reward-Based Grading

💡 The paper introduces Long Horizon Terminal Bench, a new benchmark for testing the limits of agents on long horizon terminal tasks with dense reward based grading. Existing terminal benchmarks focus on simple problems that finish within minutes and are evaluated only by their final outcome, overlooking intermediate progress and partial solutions. In contrast, Long Horizon Terminal Bench consists of 46 long horizon tasks spanning nine categories, including experiment reproduction, software engineering, multimodal analysis, interactive games, and scientific computing. Each task is further decomposed into fine grained graded subtasks, enabling dense intermediate rewards and partial credit. This design allows evaluation to capture not only whether an agent reaches the final goal, but also how far it progresses on open ended workflows. The tasks in Long Horizon Terminal Bench typically require hundreds of episodes and minutes to hours of execution, stressing long horizon planning, long context management, and iterative debugging rather than one shot problem solving. The authors evaluate 15 frontier models and find that agents consume on average 9.9 million tokens per task, with roughly 231 episodes and 85.3 minutes of execution time per run, making Long Horizon Terminal Bench more demanding than prior terminal based benchmarks. The strongest tested model achieves 15.2 percent pass rate at a partial reward threshold of 0.95 and 10.9 percent at a perfect reward threshold of 1.0, while the mean pass rate across models is 4.3 percent and 1.7 percent under the two thresholds, respectively. These results reveal headroom for improvement. The authors further analyze failure modes and error patterns, and release Long Horizon Terminal Bench to support future progress on long horizon terminal agents.

📅 Published on Jul 9

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2607.08964
• PDF: https://arxiv.org/pdf/2607.08964
• Project Page: https://zli12321.github.io/LHTB/index.html

📊 Datasets citing this paper:
• https://huggingface.co/datasets/IntelligenceLab/Long-Horizon-Terminal-Bench

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

#LongHorizonTasks #TerminalBenchmarks #DenseRewardGrading #ArtificialIntelligenceTesting #LongTermAgentEvaluation

🔥 Orca: The World is in Your Mind

💡 The paper introduces Orca, a general world foundation model that learns a unified world latent space from multimodal data, including videos and language descriptions. The model is trained using a next-state-prediction approach, which involves predicting the next state of the world given the current state, rather than just predicting the next token, frame, or action. This approach allows the model to capture both dense natural state transitions from continuous videos and sparse meaningful state transitions from language-described events.

The model is trained on a large-scale dataset consisting of 125K hours of video data and 160M event annotations. The training process involves two complementary paradigms: unconscious learning, which captures state transitions from videos, and conscious learning, which models state transitions from language descriptions and visual question answering supervision.

The results show that Orca outperforms similar-sized specialized baselines in downstream tasks, including text generation, image prediction, and embodied action generation. The model's backbone is frozen, and only the lightweight modality-specific decoders are trainable, demonstrating the scalability of the proposed paradigm. The experiments verify that a stronger world latent space enables stronger downstream readouts, indicating that Orca presents a promising approach to understanding, predicting, and acting upon the world.

The paper's contributions include the introduction of a unified world latent space, a next-state-prediction modeling approach, and a large-scale world-learning inventory dataset. The results demonstrate the effectiveness of Orca in learning a general world foundation model that can be applied to various downstream tasks, and the paper discusses the current limitations and potential future directions for the model. Overall, the paper presents a novel approach to learning a unified representation of the world and demonstrates its potential for a wide range of applications.

📅 Published on Jun 29

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2606.30534
• PDF: https://arxiv.org/pdf/2606.30534
• Project Page: https://orca-wm.github.io/

🤖 Models citing this paper:
• https://huggingface.co/BAAI/Orca-4B

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

#MultimodalLearning #WorldFoundationModels #NextStatePrediction #LatentSpaceLearning #MultimodalFusion

🔥 Signals: Trajectory Sampling and Triage for Agentic Interactions

💡 The paper proposes a signal-based framework for efficiently triaging agentic interaction trajectories in large language model applications. The problem addressed is that reviewing agent trajectories is slow and cost-prohibitive due to their voluminous and non-deterministic nature. To solve this, the authors introduce a lightweight framework that computes low-cost indicators, or signals, from live interactions and attaches them as structured attributes for trajectory triage. These signals identify informative samples without impacting online agent behavior and are organized into a taxonomy spanning interaction, execution, and environment aspects. The method is evaluated in a controlled annotation study on a widely used benchmark, where signal-based sampling achieves an 82 percent informativeness rate, outperforming heuristic filtering and random sampling. The results show a 1.52 times efficiency gain per informative trajectory, with the advantage being robust across different reward strata and task domains. The paper concludes that lightweight signals can serve as practical sampling infrastructure for agentic systems, providing a path toward preference data construction and post-deployment optimization. Overall, the framework offers a efficient and effective way to identify informative agent trajectories, enabling improved post-deployment optimization of agentic applications.

📅 Published on Apr 1

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2604.00356
• PDF: https://arxiv.org/pdf/2604.00356
• Project Page: https://planoai.dev/

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

#AgenticInteractions #TrajectorySampling #TriageForLanguageModels #EfficientTriageMethods #AgenticTrajectoryAnalysis

🔥 Arch-Router: Aligning LLM Routing with Human Preferences

💡 The paper proposes a preference-aligned routing framework called Arch-Router to improve the use of large language models. The problem with existing routing approaches is that they do not capture human preferences and typically select from a limited pool of models. To address this, the authors introduce a compact 1.5B model that learns to map queries to domain-action preferences for model routing decisions. This approach allows for matching queries to user-defined domains or action types, making routing decisions more transparent and flexible. The method supports adding new models for routing without requiring retraining or architectural modifications. The results show that Arch-Router achieves state-of-the-art results in matching queries with human preferences, outperforming top proprietary models in subjective evaluation criteria. The model is available for use, demonstrating the effectiveness of the preference-aligned routing framework in operationalizing the use of different large language models.

📅 Published on Jun 19, 2025

🔗 Links:
• GitHub: https://github.com/huggingface
• arXiv: https://arxiv.org/abs/2506.16655
• PDF: https://arxiv.org/pdf/2506.16655
• Project Page: https://huggingface.co/katanemo/Arch-Router-1.5B

🤖 Models citing this paper:
• https://huggingface.co/katanemo/Arch-Router-1.5B
• https://huggingface.co/katanemo/Arch-Router-1.5B.gguf
• https://huggingface.co/Mungert/Arch-Router-1.5B-GGUF

🚀 Spaces citing this paper:
• https://huggingface.co/spaces/build-small-hackathon/multi-agent-lab
• https://huggingface.co/spaces/tejasashinde/archRouter_simulator
• https://huggingface.co/spaces/Huggingansuman/red_teaming_env

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

#LLMRouting #HumanCenteredAI #PreferenceAlignedModeling #DomainActionMapping #TransparentModelRouting