This paper explores the concept of agentic robotics, where robots use foundation models to make decisions and perform tasks in a closed loop. Practitioners might care because agentic robotics has the potential to enable robots to work safely and reliably with increasing responsibility.
Firehose
Filtered to Papers · clear filters
Browse: People · Companies · Papers · Podcasts · Hacker News · Deep dives
Browse by tag
This paper investigates how large language models (LLMs) review papers, finding that they often focus on fine-grained technical details in the appendix, rather than the paper's core contributions, which can lead to biased reviews. Practitioners may care about this issue to ensure that AI reviews accurately reflect the paper's true strengths and weaknesses.
This paper develops a new text-to-SQL system called MEM-SQL that uses a memory to store historical SQL queries and generate new queries more accurately and efficiently. Practitioners in the field of natural language processing and database systems may care about the potential to improve the performance and accuracy of text-to-SQL systems in real-world applications.
This paper recovers vehicle trajectories from CAN signals in naturalistic driving and presents a new dataset for evaluating such methods. Practitioners might care because it shows how to accurately track vehicle movements even when GPS data is removed, which could be useful for usage-based insurance and location-based services.
This paper introduces a new AI model called Gamer that can play complex video games and interact with both virtual and physical environments. A practitioner might care about this research because it could lead to the development of more versatile and efficient AI agents that can perform a wide range of tasks.
This paper investigates how to align artificial intelligence teams, which are groups of individual models working together, and finds that certain interventions can help preserve alignment. Practitioners who design and deploy AI teams may care about understanding how to ensure their teams behave in a safe and aligned manner.
This paper introduces JVT, a simple video diffusion model that generates high-quality videos without relying on complex architectures or additional components. Practitioners might care about this paper because it explores a straightforward approach to video diffusion, which could lead to more efficient and effective video generation models.
This paper develops a neural diarization system that can handle long meetings and large archives without retraining, allowing it to accurately identify multiple speakers without a fixed number of output slots. Practitioners might care about this approach for handling large and dynamic audio recordings.
This paper creates a benchmark for large language models to evaluate their performance in finance, specifically in Japan, and shows that some models can excel in this domain even with limited external knowledge.
This paper improves the speed of generating text and images in vision-language models by adapting the way the model proposes tokens, and it could be useful to developers who want to build faster and more accurate models for applications like image captioning or text-to-image synthesis.
This paper evaluates the accuracy of underwater image quality metrics in matching human preferences, finding that none consistently agree with human judgments, and proposes a new benchmark called TrustUIQA to address this issue.
This paper develops a new type of video-language model that can read and understand the dynamics of a video, such as motion and changes in scene, rather than just relying on static text-based representations. Practitioners working on video analysis and question-answering tasks may care about this research as it could lead to more accurate and informative models.
This paper proposes a method for training conversational shopping agents in e-commerce to adapt to changing user needs and evolving product capabilities without requiring frequent retraining. Practitioners might care about how to develop agents that can learn from real-world interactions and stay effective over time.
This paper introduces a new benchmark for web agents that can perform tasks across multiple web applications and modalities, including text, vision, audio, and video. Practitioners can care about this benchmark because it provides a comprehensive evaluation framework for web agents that can handle complex, real-world tasks.
This paper develops a new AI model called DySENSE that can generate realistic energy consumption data for cities, which is essential for achieving the 2050 net-zero target. Practitioners in urban planning, energy management, and building science may care about this research as it provides a powerful tool for simulating and predicting energy usage in cities.
This paper introduces OneSteer, a controller that can track human motion while also allowing for adjustments to be made at test time. Practitioners might care because it has shown promising results in both simulated and real-world settings.
This paper introduces HALO, a system for collecting human demonstrations for whole-body mobile manipulation tasks, which aims to align human perception and action with the robot and preserve natural operator mobility. Practitioners in robotics and AI might care about this work as it could lead to more efficient and effective robot learning.
This paper introduces a framework called Evidence-Governed Evolution (EGE) to help autonomous AI agents make better decisions about which modifications to adopt. The framework provides guarantees that the chosen modifications will lead to improvements, and it can be used to ensure that agents don't degrade over time.
Distribution matching distillation (DMD) has emerged as a leading paradigm for distilling diffusion models into few-step students, where an auxiliary fake score model tracks the student's distribution and the discrepancy
This paper analyzes how coding agents behave when solving tasks, focusing on their actions and decision-making processes, to help understand their strengths and weaknesses. Practitioners might care about this research to improve their own coding agents' performance and adaptability.
A small fine-tuned parser that emits a well-formed but wrong tree fails silently: nothing downstream can tell. We ask whether the output format, a free design choice, controls how often this happens. On the PIZZA orderin
Robots handle much more than rigid objects: towels, cables, water, sand. The simulators used to train robots at scale are strongest on rigid bodies, and adding another material usually means a second engine and hand-writ
This paper develops a method for stabilizing drone videos in dynamic urban scenes by leveraging the motion of moving objects to improve registration accuracy. Practitioners might care about this because it could lead to more reliable and efficient video stabilization in real-world applications.
This paper introduces PixVid, a method that converts latent video representations into pixel space, allowing for high-resolution, controllable, and interactive video generation. Practitioners might care about PixVid because it could improve the efficiency and flexibility of video generation models.
Generalist manipulation policies can benefit from the complementary experience of humans, real robots, and simulated agents. However, differences in morphology, sensing, and control interfaces make it difficult to learn
This paper develops a way to make large language models more efficient by sharing computation across different depths, rather than using a smaller decision model. Practitioners might care because this approach can improve the accuracy and efficiency of decision-making models.
This paper creates a system called PhysLoop that helps video generation models make more realistic videos by using the knowledge that large language models have about physics. Practitioners working on video generation might care about this because it could make their models more reliable and easier to use.
This paper introduces UniSpark, a large multimodal model that uses hybrid attention to combine visual and text understanding and generation capabilities. Practitioners might care because UniSpark can perform a range of tasks, including text-image prediction, image synthesis, and video editing.
This paper explores how video models can be used to render 3D worlds created by code, and whether they can accurately preserve the world's structure and appearance over time. Practitioners might care because reliable generative rendering is crucial for applications like video games and animation.
This paper proposes a more efficient version of the SOAP optimizer, called Factored One-sided Adam-Moment (FOAM), which retains the benefits of SOAP while using less memory and running faster. Practitioners caring about optimizing large language models may benefit from this more efficient optimizer.
This paper explores the capabilities of decision models in long-horizon agents, examining whether a poorly performing model can still be effective in teaching a policy through reinforcement learning. Practitioners might care about the implications for designing more robust and effective reinforcement learning systems.
This paper introduces Ropedia Xperience 2, a large dataset of 3D scans and egocentric recordings of robots performing various tasks, which can be used to improve the performance of physical AI models by scaling their capabilities to different environments and behaviors.
This paper proposes a method to improve the rendering speed and quality of decoder-only view synthesis models by learning to incorporate global scene information into a lightweight neural network. Practitioners might care about this method if they work on rendering high-quality images quickly, as it achieves significant speedups while maintaining state-of-the-art quality.
This paper tests how well search agents can handle poor-quality evidence, which is common in real-world scenarios, and proposes a new framework to evaluate their robustness. Practitioners might care because understanding how agents perform under stress can help improve their reliability and trustworthiness.
This paper develops a new approach to reinforcement learning for large language models that incorporates both scalar rewards and fine-grained environment information, allowing the model to learn from its past experiences and improve its performance over time. Practitioners might care about this method because it has the potential to improve the performance of agentic tasks in applications where complex decision-making is required.
This paper investigates how visual input should be encoded in multimodal language models and finds that dedicated vision capacity is more important than a visual prior or pre-fusion placement. Practitioners might care because understanding how to effectively encode visual input can improve the performance of multimodal models in various applications.
This paper introduces a new approach to decision-making models that can choose whether to perform explicit reasoning or rely on implicit thinking, and how to optimize this decision based on expected benefits. Practitioners might care about this approach for tasks where explicit reasoning is costly or unnecessary.
This paper introduces a benchmark for evaluating agents' ability to generate 4D scenes in Blender, which can be used to generate real-world videos with controllable layout, motion, and camera. Practitioners can use this benchmark to assess and improve the performance of agents in this task.
This paper proposes a new framework for measuring skill acquisition in lifelong learning, where an agent can learn and perform tasks over time with feedback and rewards. Practitioners might care about this work because it aims to improve the performance of models in real-world applications where tasks are constantly changing.
This paper introduces MemInStream, a benchmark to evaluate memory capabilities in streaming video understanding, and provides a systematic way to assess memory limitations by distinguishing between understanding failures and forgetting. Practitioners may care about this research to improve their video understanding models' ability to retain and retrieve relevant information over time.
This paper introduces a new type of world model that can both perceive and interact with the environment in real-time, enabling full-duplex conversation and more intuitive human-computer interaction. Practitioners might care because this could lead to more natural and spontaneous interfaces in various applications.
This paper introduces SO-Arena, a benchmark that evaluates the capabilities and reliability of System One models, which are fast decision-making models, and shows that even the best-performing models can make unreliable decisions, highlighting the need for a more comprehensive understanding of these models' strengths and limitations.
This paper creates a publicly available framework (Open Vision Banana) to turn pre-trained image generators into general-purpose vision systems that can understand and perceive real-world images. Practitioners can use this framework to improve their own vision tasks by adapting and fine-tuning the pre-trained models.
This paper investigates the design of robot-use agents, specifically how to improve their performance on a robot arm task, and provides a simple framework for achieving significant gains without requiring complex models or additional tuning.
This paper develops a machine intelligence model called Neuro-VL that can interpret medical images and communicate its findings in a way that's useful for doctors. Practitioners might care about this because it could help improve diagnosis and treatment by providing more accurate and comprehensive information.
This paper creates a game environment called GTAGym to test AI agents' ability to complete tasks in AAA games like GTA V, and evaluates six large language models on it. Practitioners care because this work can help improve AI agents' performance in complex tasks that require reasoning and control.
This paper trains a large language model to predict molecular interactions and discovers new capabilities that go beyond its training objective, allowing it to make accurate predictions and even discover new molecules with potential therapeutic applications. Practitioners might care because this approach could revolutionize the way we design new medicines.
This paper introduces Mobile AI Factory, a system that automates the process of fine-tuning language models for mobile devices, allowing for more private and efficient AI on the go. Practitioners can care about the implications of this work for developing personalized AI models that run on mobile devices without relying on cloud connectivity.
BirdDreams: Joint Perspective and BEV Video Generation for a Plausible, Reactive Driving World Model
This paper develops a driving world model called BirdDreams that generates both a realistic video from the driver's perspective and a schematic bird's-eye-view video of the scene layout, allowing for more plausible and reactive driving simulations. Practitioners might care because this model can generate more realistic driving scenarios without requiring detailed vehicle-state inputs.
This paper proposes new methods to improve the robustness of machine learning models to adversarial attacks by directly attacking and defending against confidence structure deformation. Practitioners might care because these methods can lead to more reliable models that are less vulnerable to attacks.
This paper investigates how self-improving multimodal agents, which don't need human annotation, work across different scenarios and whether their improvements can be transferred between them. Practitioners might care about understanding how to make these agents more versatile and reliable.
This paper explores using weight-space programming to modify and train neural networks on biosignal data like EEG and ECG, allowing for "training" without gradient descent. Practitioners might care about this approach for developing more efficient and edge-friendly models for applications like wearables.
This paper proposes a method to improve the performance of large language model (LLM) agents when quantized, which reduces their computational cost, by using a full-precision teacher to guide the student to correct its mistakes and adapt to task requirements. Practitioners might care about this method because it could lead to more accurate and efficient AI agents in real-world applications.
This paper introduces PolyRoute, a benchmark for evaluating typed-decision encoders' performance across multiple languages, including code-mixed texts. Practitioners may care about PolyRoute because it provides a standardized test for multilingual language models and can help identify areas for improvement.
This paper investigates the trade-off between using fewer bits to store weights in local language models and averaging the order of answer options to reduce sensitivity. Practitioners might care because optimizing this balance can lead to more accurate and efficient models, especially in resource-constrained environments.
This paper proposes a new method to dynamically adjust the precision of weights in a small local model during decoding, and evaluates its effectiveness in improving accuracy. A practitioner might care about this approach if they want to improve the performance of their small local models, especially in scenarios where the output distribution is uncertain.
This paper develops a faster GPU-accelerated solver for solving quadratic programs used in real-time model predictive control, which is important for applications like robotics and autonomous systems where fast decision-making is crucial.
This paper creates a benchmark to evaluate how well AI agents follow website scraping rules and laws, and finds that many AI agents don't follow these rules even when they're supposed to. Practitioners might care because they want to ensure their AI systems don't accidentally break the law.
This paper develops a more efficient method for generating videos on demand by reducing the need for extensive training data. Practitioners might care about this research because it could lead to faster and more cost-effective video generation for applications like video editing and content creation.
This paper introduces a method to adapt long action chunks in robot policies, allowing for more efficient online learning with fewer samples. Practitioners might care about this approach for developing capable robot policies that can adapt to new situations with minimal data collection.