Filtered to continual learning · clear filter
This paper proposes a method to scale automatic research agents by replacing environment execution with a world model, which can reduce training costs and improve performance. Practitioners might care about this approach because it can accelerate training times and lead to better results for complex AI tasks.
This paper introduces Atria Dawn Preview, a new type of AI model designed to work alongside humans in scientific research and engineering. Practitioners might care about this because it shows how AI can collaborate with humans more effectively, potentially leading to better research outcomes and more autonomous AI development.
This paper develops a method to create personalized AI tutors that can adapt to individual students' strengths and weaknesses, using a combination of pooled training and per-student fine-tuning. Practitioners in education and AI development may care about this work as it could lead to more effective and personalized learning experiences.
This paper introduces a method to compile neural functions from natural-language specifications, allowing for faster and more reliable execution without relying on remote models. Practitioners may care about this approach for building efficient and flexible AI systems that can perform complex tasks without the need for expensive model calls.
This paper develops a vision-language model for autonomous driving that combines 3D perception, question answering, and motion planning. A practitioner might care because it demonstrates a promising approach to integrating multiple tasks in autonomous driving.
This paper proposes a method for recursive self-improvement in AI systems, where a model can learn from its own performance and use that knowledge to improve itself. Practitioners might care about this approach because it could lead to more efficient and effective AI systems.
This paper introduces a new type of language model called NCP-ArchPreview that can generate text by predicting both individual tokens and larger concepts, and how this approach can improve performance. Practitioners might care about this because it could lead to better language models that can handle more complex tasks.
This paper introduces ZGCM-1, a highly efficient foundation model for math and agentic search that combines internal thinking with external tool use, and shows it can perform well on various benchmarks despite its compact size. Practitioners may care about the efficiency improvements and scalable architecture of ZGCM-1.
This paper explores a way for language models to remember information from previous updates, even after many changes, and finds that combining different techniques can lead to much better retention than using any one technique alone.
This paper introduces Qwen-UI-Agent, a type of artificial intelligence system that can perform tasks on various devices, such as smartphones and computers, and improve its abilities on its own. Practitioners might care about this research because it aims to create more practical and autonomous AI systems that can be used in real-world scenarios.
This paper develops a method to turn agent trajectories into reusable environments, allowing for more efficient testing and interaction with the agent. Practitioners might care because this approach can help scale agent training and improve performance in various applications.
This paper introduces a new type of AI model called memory foundation models, which allows the model to learn and retain information internally, rather than relying on external memory modules. This could be useful for practitioners who want to build more efficient and flexible AI agents.
This paper introduces OpenART, a platform for testing AI agents in open-ended environments, to evaluate their safety in complex and evolving scenarios. Practitioners can use OpenART to identify potential risks and improve the robustness of their AI systems.
This paper creates a benchmark (HarnessDev) to test whether large language models (LLMs) can design and improve their own execution infrastructure, called the agent harness, which affects their performance. Practitioners might care because it explores how models can adapt to changing environments and potentially improve efficiency.
This paper introduces ASPIRE, a benchmark for self-evolving models that can learn from vague goals, and explores how agents can interpret and operationalize these goals to improve their performance. Practitioners may care about how models can learn from unclear objectives in real-world applications.
This paper shows how to make large language models improve themselves without needing external guidance or supervision, by using their own internal consistency to correct mistakes. Practitioners might care about this because it could lead to more robust and self-sufficient AI models.
This paper introduces VGI-bench, a new benchmark for evaluating video generation models' visual reasoning capabilities, and finds that current models can solve some visually grounded tasks but still struggle with reliability. Practitioners may care about developing more reliable video generation models that can perform better on this benchmark.
This paper introduces Macaron-V1, an open agent model family that enables continual learning and self-improvement in real-world environments, and explores its potential for collective intelligence. Practitioners might care about this work if they're interested in developing AI systems that can learn and adapt over time.
This paper develops a new approach to understanding videos by predicting when specific events or evidence occur within the video. Practitioners working on video analysis and AI models might care about this research because it could lead to more accurate and robust video understanding systems.
This paper investigates how skills work in LLM agents, when they help, and why they fail, to improve their performance. Practitioners can benefit from understanding these factors to build more reliable self-evolving agents.
This paper proposes a method to determine which past update evidence in a large language model is still relevant and useful after subsequent training, to prevent wasting compute and potentially degrading the model's performance. Practitioners in the field of autonomous systems and language models might care about this problem because it can lead to better model performance and efficiency in adapting to changing domains and requirements.
This paper introduces SolarWM, an open framework for training video world models from diverse datasets and using different video backbones, allowing for more consistent and reproducible results. Practitioners can use SolarWM to build interactive video world models that can be applied to various real-world applications.
This paper introduces RAGU, an open-source GraphRAG engine that improves large language models with structured knowledge by separating extraction and consolidation, and trains a compact extractor that outperforms larger models on knowledge-graph construction and GraphRAG tasks. Practitioners might care because RAGU can efficiently generate more accurate and complete context for tasks like factoid-level evidence recall and multi-hop question answering.
This paper introduces FrontierChallenge, a benchmark to evaluate the completion of end-to-end scientific workflows, and investigates the performance of various models in completing these workflows. Practitioners might care about this research to improve the accuracy of scientific workflow completion evaluations.
This paper explores how agentic systems can improve on their own through co-evolution, where multiple agents and their environment adapt to each other, and discusses the challenges and benefits of building such systems that can learn beyond human design.
This paper creates a platform to help large language models generate code for data pipelines, which can then be edited and used to automate data processing workflows. Practitioners might care about this because it can help reduce the time and cost of developing and maintaining these pipelines.
This paper proposes a framework, LongHorizon-Harness, to help large language model agents tackle long-horizon tasks by explicitly tracking task states and verifying facts from the environment. Practitioners might care because it can improve the performance of these agents on real-world tasks.
This paper introduces TurboVLA, a new vision-language-action model that reduces computation and memory overhead by directly exchanging information between visual observations and language instructions, allowing for faster and more efficient robotic manipulation. Practitioners might care about this approach for building more efficient and effective VLA models.
This paper develops a method to generate high-quality terminal tasks for training agents, ensuring that the tasks accurately reflect the original instruction and environment, and providing a way to validate and improve the generated tasks. Practitioners in AI and robotics may care about this work because it addresses a common challenge in training terminal agents.
This paper develops a system for generating speech and audio for various applications, including animation and video production, without reference recordings. Practitioners can use this system to create customized voices and control speaker styles, and to generate high-quality audio in complex scenarios.
This paper develops a new AI model called EVOKE that can generate interactive worlds without running out of memory or taking too long. Practitioners might care about EVOKE because it can support long-term generation and adaptation in complex environments.
This paper explores how AI can be applied across different stages of game development, from playing games to designing and testing them, and how to reuse capabilities across these stages. Practitioners might care about how to apply AI to improve game development efficiency and effectiveness.
This paper introduces EarlyEval, a method to reduce the cost of evaluating large language model (LLM) agents by predicting their outcomes early, allowing for earlier termination of agent runs. Practitioners may care because it can significantly reduce the computational cost of agent development.
This paper introduces LatentPress, a method to compress conversational histories and documents into a continuous memory token format that allows language models to directly read and process the context without text reconstruction. Practitioners might care about this because it could lead to faster and more efficient language model inference.
This paper develops a model that can automatically generate and adapt agent harnesses, which are crucial for the performance of AI models, to improve their ability to perform tasks. Practitioners might care about this because it could lead to more efficient and effective AI systems.
This paper explores how large language models can turn local observations of a city into reliable actions, and whether these models can sustain goal-directed behavior in complex urban environments. Practitioners in AI/ML and urban planning might care about the limitations and potential of current models in navigating real-world cities.
This paper presents a model called Gander that can understand and interact with users in multiple ways, like talking, typing, and watching videos, in real-time. A practitioner might care about this research because it could lead to more natural and expressive human-AI interactions.
This paper introduces StepAudio 3 Realtime, a model that enables real-time spoken interaction with deep reasoning, prompt responses, and fluid turn-taking. Practitioners might care because this technology can improve human-computer interaction, enabling more natural and efficient conversations.
This paper introduces LoopArena, a benchmark for evaluating how well a model can guide a coding agent through a long-running task, and finds room for improvement in long-horizon loop control. Practitioners working on AI development might care because it could lead to more efficient and reliable development processes.
This paper presents GigaBrain-0.7, a new embodied foundation model that achieves strong generalization across diverse robot embodiments and tasks, by improving the architecture and scaling it to large amounts of data. Practitioners might care about this research if they're working on developing generalist robots that can adapt to new tasks and environments.
This paper proposes a method to optimize prompts for multi-agent systems by identifying which agent's modification resolves a failure, and then using that agent's output as supervision to extract a fine-grained gradient. Practitioners might care because it could improve the performance of large language model-based multi-agent systems.
This paper proposes a new method for fine-tuning large language models (LLMs) in reinforcement learning (RL) tasks with long horizons, using evolution strategies (ES) instead of traditional backpropagation-based training. Practitioners might care because it allows for more efficient and flexible fine-tuning of LLMs, enabling them to tackle complex tasks with larger models and longer interactions.
This paper improves autoregressive video distillation methods by aligning the initialization and distribution matching stages, focusing on matching the target distribution's mode coverage rather than just visual quality. Practitioners can benefit from this approach to generate higher-quality videos with better diversity and coverage.
This paper investigates whether on-policy distillation (OPD) truly improves student policies by analyzing the effects of noisy teacher supervision. It finds that OPD works by suppressing low-probability tokens, which can be achieved without a teacher, and introduces a new method called On-Policy Self-Adaptation (OPSA) that outperforms OPD and traditional reinforcement learning methods.
This paper proposes a new framework for training autonomous agents to perform multi-turn tool-calling tasks, addressing the challenge of dealing with vast solution spaces by using a diamond topology-aware approach. Practitioners may care about this research because it can help improve the performance of autonomous agents in complex tasks.
This paper introduces a framework called DeepSearch-Evolve, which helps train self-improving web agents by iteratively refining their performance using their own experience. Practitioners might care because this approach can lead to more efficient and effective agents that can learn from their own mistakes.
This paper introduces a new technique called SkillZip that helps compress large libraries of skills used by artificial agents, making them more efficient and scalable. Practitioners who work on developing agents that can learn and adapt in complex environments might care about SkillZip because it can improve the performance and efficiency of their agents.
This paper presents a data-centric framework to overcome challenges in training cyber agents, allowing for more efficient and effective training of open-weight models, and demonstrates the capability of a team of seven to train such models with leading agentic cyber capability.
This paper proposes a method to improve language model training by allocating credit to individual tokens within a response, allowing for more nuanced evaluation of model performance. Practitioners may care about this method as it can lead to better language model performance, especially in tasks that require specific formatting or semantic choices.
This paper introduces a new method for training flow matching models called Self-OPD, which uses the model's own exploration to generate supervisory signals without needing a separate teacher model. Practitioners may care because it could lead to more efficient and effective training of flow matching models.
This paper investigates why a specific neural network architecture, Gated DeltaNet, can survive 4-bit quantization. The researchers found that a new quantization technique, NVFP4, can be used to quantize the recurrent state of GDN without significant loss of performance, and they provide a mechanistic explanation for why this is the case.
This paper introduces DarwinX, a method for evolving agent capabilities through natural selection, which improves the agent's overall performance on various benchmarks without relying on task-specific patches. Practitioners might care about this approach as it allows for the creation of more general and durable agents that can adapt to changing environments.
This paper proposes a new method called Test-Time Policy Optimization (TTPO) that allows large language models to be trained without ground-truth labels, enabling test-time training. Practitioners may care about this because it enables models to be trained without labels, which can be difficult or expensive to obtain.
This paper investigates how to improve the adaptation of diffusion models in a way that doesn't rely on a classifier, and how to address a problem where the model can't accurately learn from its teacher. Practitioners might care about this because it could lead to more effective knowledge transfer in machine learning applications.
This paper develops a standardized evaluation method for computer-using agents (CUAs) to ensure they fulfill task instructions, using vision-language models (VLMs) as judges. Practitioners can benefit from this work by using reliable and cost-effective reward signals for CUA training.
This paper introduces Ouroboros, a self-improving AI agent that develops its own tools and code through a process of reviewed commits, allowing it to learn and adapt over time. A practitioner might care about Ouroboros because it demonstrates a potential approach to creating autonomous AI systems that can improve themselves.
This paper proposes a new pruning method for coding agents that prunes tool outputs directly inside the agent, rather than relying on a separate code classifier, and shows it can save up to 39% of tokens while preserving task quality.
This paper proposes a new approach to Large Language Models (LLMs) that uses diffusion to speed up generation without sacrificing quality, allowing for faster and more efficient language processing. Practitioners may care about this research if they need to process large amounts of text quickly and efficiently.
This paper introduces Grouped Value Attention (GVA), a technique to reduce the memory footprint of Transformer decoding by storing grouped values and reconstructing content keys. Practitioners might care about this because it could lead to faster and more efficient models.
This paper proposes a new retrieval framework called CoGR, which uses two generative models to co-evolve and optimize retrieval representations on both the query and item sides, leading to improved search and advertising performance. Practitioners might care because CoGR can potentially lead to better retrieval results and more efficient search systems.
This paper proposes a method called Latent Interface Training (LIT) to improve the generalization of robotics foundation models by preventing them from relying on visual shortcuts when learning to generate actions from pre-trained visual representations. This is important for robots to perform well in new, unseen environments.
This paper introduces SWE-bench Science, a benchmark for evaluating coding agents' performance in repairing scientific software, and identifies common failure mechanisms that hinder their success. Practitioners may care about this research as it can help improve the reliability and reproducibility of scientific findings by developing more effective coding agents.
This paper develops a framework for generating agentic data that is consistent, useful, and sufficient for LLM agents to learn from. Practitioners might care because they want to ensure their agents learn from high-quality data that helps them adapt to changing environments.
This paper introduces Mage-Flow, a compact model for generating and editing high-resolution images, which can be trained efficiently and deployed on a single GPU. Practitioners might care about the potential applications of this model in interactive image editing and generation tasks.
This paper introduces WeMM-Embedding, a family of universal multimodal embedding models that can represent various types of data in a shared space, enabling applications like retrieval, recommendation, and classification. Practitioners might care because WeMM-Embedding achieves state-of-the-art performance on multiple benchmarks and has been deployed at scale in real-world applications.
This paper introduces a vision-language-action model that can perform mobile manipulation tasks in unseen environments with minimal training data, and how it can be scaled up to achieve better performance. Practitioners might care about this model for building robots that can adapt to new tasks with minimal fine-tuning.
This paper develops a framework to create customizable safety harnesses for AI agents, which can adapt to different models and domains to prevent harmful behavior. Practitioners may care about this research to improve the safety of AI systems in real-world applications.
This paper develops a new AI model that can efficiently search through vast spaces of possibilities, such as designing molecules or optimizing neural networks, by combining a generative model with a model that estimates the potential performance of each candidate. Practitioners might care about this approach because it can significantly speed up the discovery process and improve the quality of the results.
This paper introduces a new approach to training latent generative models, called GenFirst, which trains a generative model before reconstruction, and shows that this approach can achieve stable end-to-end training and state-of-the-art results on image generation tasks.
This paper introduces ASI-Bench, a new benchmark for evaluating AI systems' ability to explore new knowledge, create new ideas, and conduct scientific research autonomously. Practitioners might care about ASI-Bench because it aims to push the limits of current AI systems and accelerate the development of artificial superintelligence.
This paper proposes a new approach to training Vision-Language-Action models by using a pre-trained backbone that captures generalizable visual-action knowledge from a large, diverse dataset of robot trajectories. This allows the model to perform well on new, unseen tasks without requiring a large amount of task-specific data.
This paper introduces a method called DecoEvo, which helps large language models improve by co-evolving a solver skill and a rubric-generator skill in a way that's more efficient and effective. Practitioners might care about this because it could lead to better performance and more reliable optimization in open-ended tasks.
This paper presents two search agents, Iris-mini and Iris-pro, trained to solve complex search tasks using reinforcement learning and self-supervised learning. Practitioners might care because these models achieve state-of-the-art results on various benchmarks, demonstrating the potential of AI-powered search agents in real-world applications.
This paper develops a method to scale up language models by increasing their memory capacity, allowing for better performance and more efficient use of parameters. Practitioners may care about this research as it could lead to more powerful and efficient language models for applications like language translation and text generation.
This paper proposes a framework for training agents to create 3D open worlds based on user queries, and evaluates its performance using a large benchmark dataset. Practitioners might care about this research because it can help develop more capable multimodal agents that can understand user intent and generate realistic 3D environments.
This paper proposes a new method to sparsify attention in Transformers, called Simple Attention Sparsification (SAS), which optimizes context ranking end-to-end with the language modeling loss. Practitioners might care because SAS can improve performance on downstream tasks by using attention budgets more effectively.
This paper introduces a benchmark to measure how well large language models (LLMs) can prepare training data, and how well they evaluate the quality of that data. Practitioners might care because improving data preparation can lead to better model performance.
This paper introduces a vision-tactile-language-action model that can perform fine-grained manipulation with tactile perception and control, and improve its policy offline from stored data. Practitioners may care because this model can be used to create more versatile and accurate tactile-driven manipulation policies.
This paper introduces a new method for improving diffusion models by using self-distillation to align them with human preferences and task-specific objectives. Practitioners might care about this approach because it can lead to more efficient and analyzable alignment of diffusion models with human goals.
This paper trains compact AI models to adapt quickly to changing digital avatar "harnesses" that define the tasks and tools available to the model, improving performance and reducing latency. Practitioners caring about real-time AI applications, such as chatbots or virtual assistants, might find this approach useful.
This paper proposes a new method for estimating confidence in language models, called XConf, which uses the model's past experiences to inform its confidence, rather than just relying on the current inference process. Practitioners might care about this because it could lead to more reliable and trustworthy deployment of language models.
This paper introduces Agent Memory Distillation, a technique that allows small language models to learn from a larger teacher model by transferring structured knowledge through hierarchical memory. Practitioners might care about this approach because it could improve the performance of small language models in tasks that require complex decision-making.
This paper proposes a new framework for embodied navigation that overcomes limitations of existing methods by reformulating navigation into a 2D visual space, introducing selective reasoning and memory mechanisms, and designing an efficient alignment paradigm. Practitioners caring about efficient navigation for AI agents might care about this research.