Firehose

This paper proposes a method to improve dense 3D tracking from monocular video by treating it as a geometric trajectory repair and completion task. Practitioners might care about this research because it provides a more efficient and accurate approach to 3D tracking, especially when using pre-trained trackers and geometry models.

This paper proposes a runtime system called REA that enables the coordination of fast, reactive policies (like those used in vision-language models) with slower, deliberative brain models, allowing them to work together effectively. Practitioners might care about REA because it can improve the performance and efficiency of embodied systems like robots.

Existing sketch generation methods often rely on image guidance and are largely limited to individual objects or simplified scenes, leaving text-guided complex scene sketching underexplored. We present SketchDirector, an

This paper develops a method for short-video search that uses human feedback to improve the accuracy of query rewriting, taking into account the user's immediate needs rather than just their long-term preferences. Practitioners might care about this because it can lead to better search results and user engagement.

World action models (WAMs) learn latent variables from observations and use these to both predict the evolution of the scene and to generate actions. The prediction has to be of the whole scene, whereas only a portion of

This paper develops a system called RoboArchitect that automates the design of robotic agents by learning to optimize their harnesses, which determine when the agent interacts with the environment. A practitioner might care about this because it could significantly improve the performance of robotic agents without requiring significant updates to their underlying models.

This paper develops a new framework called Latent Reasoning in Continuous space (LARC) that helps Unified Multimodal Models better use language to generate images, leading to improved image quality and controllability. Practitioners might care about this research if they want to create models that can generate images that are more coherent and flexible.

This paper presents TurnTurk-GPT, a text-based turn-taking predictor for Turkish obtained by fine-tuning a Turkish GPT-2 model with the TurnGPT objective on synthetic and real dialogue, and evaluates it on real Turkish c

Federated learning (FL) keeps raw data local, but most deployments still rely on a trusted coordinator for client admission, checkpoint distribution, update collection, aggregation, and rewards. This paper presents Audit

This paper investigates how training acceleration in diffusion transformers works and why it's important for practitioners to understand this process. By analyzing how the residual gate is reorganized during training, the authors found that this reorganization is a key factor in accelerating training, and they propose a simple method to promote this reorganization.

This paper investigates whether visual foundation models can support spatial state representations that are updated under self-motion and used to guide action, which is crucial for tasks like spatial perception and object tracking. Practitioners working on visual perception, object tracking, or action planning might care about this research.

Do research agents have taste, the judgment of which experiments are worth running before their outcomes are known? We give six frontier models seven recent NLP papers. Each model, working as a research agent alone or as

This paper creates a voice agent tool called VoxLock that can automatically identify and transcribe the speaker's voice without needing to enroll them beforehand, which is useful for applications like customer service. Practitioners might care about VoxLock because it can improve the accuracy and efficiency of voice-based interactions.

This paper creates a system that turns users' histories into labels to help improve AI agents' memory systems, making them more adaptable to different tasks and users. Practitioners might care about this because it can help improve the performance of AI agents in real-world applications.

This paper develops a new framework called HiFOR that helps vision-language-action models predict and plan for future events in a more accurate way, which can improve their performance in tasks like robotic manipulation. Practitioners can care about this research because it could lead to more reliable and effective robots that can adapt to changing situations.

This paper develops a method to improve robots' ability to generalize to new tasks by using both video and action predictions. By combining these predictions, the robot can better adapt to new situations, leading to better performance on a variety of tasks.

This paper explores whether achieving artificial general intelligence (AGI) depends on improving a lightweight predictor (System-1) rather than just scaling a resource-intensive executive (System-2). Practitioners might care because understanding how to design and optimize these systems could lead to more efficient and effective AI.

This paper introduces Wizard, a system that lets players control and customize virtual worlds in real-time, generalizing to new environments and characters without sacrificing visual quality or gameplay consistency. Practitioners in game development and virtual reality might care about this technology for creating more immersive and interactive experiences.

This paper introduces CaPtain, a post-training pipeline that enables small language models to become capable robot agents that can write code against robot APIs and read back the results. A practitioner might care about this work because it demonstrates a novel approach to training robot agents using language models, which could be applied to various robotics tasks.

This paper investigates the problem of training-inference mismatch in reinforcement learning for large language models, where the training and inference processes can produce different results, leading to suboptimal training. A practitioner might care about this issue because it can lead to poor performance and instability in training, especially when using off-policy methods.

This paper introduces Apriel Voice 27B Preview, an open speech-language model that combines capabilities for understanding requests, reasoning, and acting through tools. Practitioners caring about developing conversational AI may be interested in Apriel Voice as it offers a unified approach to speech, reasoning, and tool use.

This paper develops a new method to evaluate proactive agents, which are systems that decide when and how to assist without being asked. Practitioners in AI and robotics might care because it helps them understand how well these systems work and make them more effective.

This paper develops a framework called CodeSpecWeave to help coding agents create more realistic and comprehensive task specifications for long-horizon software development. Practitioners may care about this because it could improve the performance and reliability of coding agents in real-world development scenarios.

This paper introduces EditAgent, a new approach to agentic image editing that allows for more flexible and reliable editing by combining general-purpose and specialized methods, and verifying candidate edits with off-the-shelf judges. Practitioners may care because it can improve the efficiency and quality of image editing tasks.

This paper proposes a method to improve text-to-image generation models by matching the generated distribution with a reference distribution, allowing for more accurate rendering of details specified in the prompt. Practitioners may care about this approach as it can lead to better performance on image generation tasks.

This paper proposes a method to generate stylized hair cards from a single image, which can be animated and exported to downstream formats. Practitioners in the field of computer graphics and animation may care about this research as it can simplify the creation of stylized characters.

This paper shows that the learning rate schedule used in weight-normalized training is more important than the normalization itself, and proposes a new schedule called AIR that targets the effective learning rate trajectory. Practitioners might care about this because it could help them improve the performance of their models by optimizing their learning rate schedules.

This paper introduces Glance, a model that combines generation and reranking into a single model, which can improve performance and reduce memory usage. Practitioners might care about this because it could lead to more efficient and effective AI-powered question-answering systems.

This paper develops an online reinforcement learning framework that trains agents to navigate UIs on live iOS apps, addressing limitations of existing models. Practitioners might care about how this approach can improve the performance and generalizability of UI navigation models in real-world applications.

This paper proposes a method for generating infinite, interactive 3D worlds in real-time, using a compact 3D memory that stores scene information efficiently. Practitioners might care about this research if they need to generate realistic, interactive environments for applications like virtual reality or video games.

This paper introduces a new architecture for generative transformers that allows them to capture finer details without increasing the length of the global attention sequence, making them more efficient and effective. Practitioners may care about this approach as it could lead to significant improvements in image and text generation models.

This paper proposes a new attention mechanism for video diffusion transformers that reduces the computational cost of inference by selectively computing attention only on key/value blocks. It improves the accuracy of attention at high sparsity levels, leading to faster inference times and better video quality.

This paper introduces a new method for inserting objects into scenes in a physically plausible way, allowing for more realistic lighting, pose, and perspective. Practitioners might care because this could be useful for applications like virtual try-on, augmented reality, or computer-generated imagery.

This paper presents a unified AI model that enables aerial robots to understand environments, navigate, and control themselves using a shared vision-language representation. Practitioners might care about how to build AI systems that can perform multiple tasks, such as perception and control, without relying on external resources.

This paper introduces Youtu-Parsing-Omni, a system that can parse documents, images, audio, and video into structured records, allowing agents to access and cite evidence across multiple modalities. Practitioners might care about this research because it provides a new approach to multimodal parsing that can improve the performance of agents in various applications.

Formula-Driven Supervised Learning (FDSL) enables vision pre-training without real images, but existing methods cast pre-training as classification on discretized labels, discarding continuity in the generative process.

This paper proves the Platonic Representation Hypothesis by showing that models trained on different modalities (e.g. vision and language) converge to the same representation of the world as they scale, but with a specific condition (good views). Practitioners might care because this has implications for building more efficient and accurate multimodal AI models.

This paper shows that foundation models can be trained with a single algorithm, which can improve the accuracy-compute tradeoff and task coverage. Practitioners might care because this approach can simplify and streamline the training process for foundation models.

This paper shows how to make existing autoregressive image models generate images in any order, without having to retrain them from scratch, which could improve their flexibility and usefulness for tasks like inpainting and editing. Practitioners might care about this because it could make their models more versatile and efficient.

This paper introduces a benchmark for evaluating the ability of large language model (LLM) agents to forecast earnings based on financial data. Practitioners in finance and AI research may care about this work because it could lead to more accurate and reliable earnings forecasting models.

This paper introduces a new method called Weakening-Induced On-Policy Self-Distillation that helps improve reasoning models without needing a strong teacher model. Practitioners might care because it can be used to improve model performance even when a good teacher model is not available.

This paper introduces a new dataset and benchmark for multimodal interleaved generation, which is the process of jointly generating text and images that are naturally intertwined like in real-world content. Practitioners might care about this because it could lead to better AI models that can create more realistic and engaging multimodal content.

This paper develops a single AI agent that can edit long-form videos from narration, improving the efficiency and quality of video production. Practitioners in the video production industry may care about this research as it has the potential to automate video editing tasks and reduce production costs.

This paper proposes a new approach to proactive interaction in video understanding, where models can ask for missing information to improve their performance. Practitioners caring about improving the accuracy of video understanding models may find this research useful.

This paper introduces A-RSI, a method that helps AI models decide when to stop learning from their current experience and introduce new tasks, without relying on external validation or human feedback. By monitoring the model's learning signal and accuracy, A-RSI can automatically refresh the task pool, leading to improved performance.

This paper introduces a benchmark, called Investor's Last Exam (ILE), to evaluate the ability of AI agents to make informed investment decisions without having access to future information. Practitioners in the field of artificial intelligence, finance, and investment management may care about this research as it aims to improve the performance of financial agents in making investment judgments.

This paper improves a common technique in language model training called Multi-Token Prediction (MTP) by removing a noisy component that interferes with learning, resulting in faster and more accurate predictions. Practitioners may care because this improvement can lead to better performance in applications such as language translation and text generation.

This paper introduces a Lean theorem proving system called EasyProver, which can efficiently verify code and solve complex mathematical problems. Practitioners might care because it could lead to more reliable and efficient code verification in real-world applications.

This paper introduces SciGame, a controlled benchmark for evaluating language model agents in scientific discovery, and finds that familiar priors can hinder performance, adaptive experimentation can be beneficial but also unfair, and that agents struggle to induce exact symbolic laws in unfamiliar settings.

This paper proposes a new method for unified image generation that leverages pre-trained video models to better understand the relationship between the input image and the desired output. By modeling input-adaptive virtual temporal relations, FrameForge can generate high-quality images that are tailored to specific tasks and applications.

This paper introduces a new AI model called Venus-DefakerOne2 that can detect a wide range of fake images, videos, and documents, and is useful for developers who need to build robust anti-forgery systems.

This paper develops a robot policy that allows for real-time control over quality attributes like speed and smoothness, and shows that this approach can improve both task success and quality in real-world deployment scenarios.

This paper presents a new approach to building large language models by reducing the parameter count while maintaining performance, and analyzes what went wrong with the approach, providing valuable insights for improving future models. A practitioner might care about this paper if they're interested in building efficient and effective language models.

This paper helps prevent large language models from becoming unsafe by identifying training data that causes them to behave badly, even if the data itself looks harmless. Practitioners can use this technique to build safer language models that are more reliable and trustworthy.

Large Language Models (LLMs) are increasingly used in multi-turn agentic workloads that require sustained interaction with complex environments. Existing reflective prompt optimizers such as GEPA learn from failed rollou

This paper proposes a new learning paradigm for visual generation that combines autoregressive pretraining for knowledge acquisition and diffusion post-training for high-fidelity rendering, aiming to improve the efficiency and quality of visual generation models. Practitioners might care about this approach because it could lead to more efficient and effective visual generation models for applications such as image synthesis and rendering.

This paper introduces WorkFactory, a framework for creating executable tasks for training agents, which can help improve the reliability of expert demonstrations and scale agent training. Practitioners might care about how to create coherent and executable tasks for their agents, especially in complex environments.

This paper introduces WebAnything, a framework that automatically generates reusable website environments from real URLs, allowing for reproducible training and evaluation of web agents. Practitioners might care about this because it simplifies the creation of diverse and realistic environments for training AI models.

This paper proposes a new method to process images only when necessary in diffusion vision-language models, reducing unnecessary computations and maintaining high accuracy. Practitioners caring about efficient model performance might be interested in this approach.

This paper introduces a new deep learning framework called Quaternion Approximate Networks (QUAN) that uses quaternions to improve image classification and object detection. Practitioners might care about QUAN because it can achieve higher accuracy with fewer parameters and faster convergence, making it suitable for deployment in resource-constrained robotic systems.