Filtered to reinforcement 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 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 Kimi K3, a large-scale, open-source AI model that achieves state-of-the-art performance on a range of tasks, including vision and coding, and is designed to be more efficient and scalable than previous models, making it a promising candidate for real-world applications.
This paper introduces EnvHarness, a tool that lets AI agents learn from static environments by changing how they interact with those environments, without having to rebuild the environments from scratch. This can help improve the performance of AI agents in real-world applications.
This paper introduces Dream-RSI, a framework for recursive self-improvement in exploration, which helps autonomous AI agents discover high-value solutions more efficiently by using a replay simulator to provide low-cost feedback. Practitioners might care because effective exploration is crucial for AI progress, and Dream-RSI can improve discovery quality and reduce costs.
This paper introduces a method to generate long-horizon training tasks for terminal agents at scale, which can be used to improve AI models' performance on tasks like navigation and decision-making. Practitioners might care about this because it could help train more advanced AI models with better performance on complex tasks.
This paper proposes a way to improve the efficiency of training world models by using game development as a source of reward signals and trajectory data, allowing for more effective post-training of large language models using reinforcement learning. Practitioners might care about this approach because it could lead to more scalable and effective world models for applications like dialogue systems and visual question answering.
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 proposes a new method called Self-Supervised Visual On-Policy Distillation (S^2VOPD) that generates learning signals from asymmetric augmented views of images, allowing for effective on-policy learning without privileged information. Practitioners might care about this paper because it presents a simple yet effective way to improve performance on various perception benchmarks.
This paper introduces AREX, a self-improving agent that recursively refines its answers by verifying intermediate results and using the partially verified state to guide subsequent refinement, aiming to improve the efficiency of deep research. Practitioners might care about AREX's approach to reducing the cost of discovery and verification in complex research tasks.
This paper improves reinforcement learning for post-training language model reasoners by addressing two problems in existing methods: identical advantages for distinct reward profiles and fixed relative weights for all objectives. A new method, SA-MRPO, dynamically reallocates optimization effort toward under-optimized objectives while maintaining performance on well-satisfied objectives.
This paper introduces Zetta, a system that enables embodied agents to learn and adapt in real-time while executing physical tasks, allowing for more efficient and effective learning. Practitioners in robotics and AI may care about Zetta's approach as it could lead to more reliable and scalable physical intelligence.
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 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 DataFlex-RL, an evaluation platform for comparing different data policies in reinforcement learning with verifiable rewards (RLVR), and finds that uniform data policy leads to better performance and is more reproducible than other methods. Practitioners might care about this because it provides a way to compare different data policies and ensure that their RLVR models are performing well.
This paper explores a new approach to transferring capabilities from strong AI models to weaker ones at test time, rather than just during training. Practitioners might care about this because it could lead to more efficient and effective use of powerful models in real-world applications.
This paper introduces a method called OraRL to improve the efficiency and scalability of reinforcement learning for multimodal large language models (MLLMs) trained on video data. By leveraging annotations as a source of high-quality rollouts, OraRL can significantly reduce the number of required rollouts, leading to faster training times and better performance on video understanding tasks.
This paper develops a new framework for world modeling that takes into account the mental state of agents, which is essential for predicting human decisions. Practitioners caring about human decision-making and planning might find this research useful.
This paper introduces Co-RL, a framework for unsupervised multi-agent reinforcement learning that enables diverse and accurate reasoning in language and vision-language models. Practitioners can use Co-RL to improve their models' ability to reason and respond without relying on expensive ground-truth supervision.
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 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 shows how a vision-language model (VLM) can control robots without needing extensive pretraining or specialized hardware, by providing a compact interface that links the model's intentions to specific actions. Practitioners might care about this because it could make robots more accessible and user-friendly.
This paper introduces a new AI model called DreamX-Phi 1.0 that can predict what will happen in a robotic manipulation scenario, given an initial state and instructions. This model is useful for robotics developers because it can help them design more reliable and efficient robotic systems.
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 method called On-Policy Reverse Distillation that helps stronger models learn from weaker supervisors by selectively amplifying the parts of the weaker model's guidance that are most useful to the stronger model. This can lead to faster and more efficient learning, especially in situations where it's expensive to retrain models from scratch.
This paper introduces a large-scale music generation model called StepAudio 3 Music, which can create long-form music and respond to text prompts. Practitioners might care about this model's ability to generate music with structure and coherence.
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 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 introduces a new framework called RSIAgent that helps digital agents adapt to new environments without needing to be retrained. A practitioner might care about this because it allows for more efficient and effective AI systems that can learn and improve on their own.
This paper introduces a new method to improve large language models (LLMs) called RLSVR, which uses a task-transformation technique to generate self-verifiable rewards, enabling LLM self-improvement on open-ended tasks. Practitioners might care about this because it could lead to more reliable and scalable self-improvement methods for LLMs.
This paper creates a system called ScienceIDE that converts scientific code into environments that can be used to train artificial agents to perform scientific tasks. Practitioners might care because this could lead to more efficient and effective ways to develop scientific intelligence.
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 proposes a new method to improve the performance of large language models on knowledge-intensive tasks by distilling knowledge from proprietary models and using reinforcement learning. Practitioners may care about this approach because it can help bridge the gap between proprietary and open-source models, leading to more effective and robust AI systems.
This paper proposes a new method for training long-horizon search agents that can search, retrieve, and integrate evidence to reach a final answer. Practitioners in natural language processing and AI research might care about this paper because it shows a way to improve the performance of search agents, which can be used in applications such as question-answering systems.
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 investigates a common problem in reinforcement learning for language models called Value Flattening, where critics fail to accurately estimate state values, and proposes a new method, SP^3O, to mitigate this issue by supervising only a few well-separated states per response.
This paper introduces UI-Venus-2, a general-purpose GUI agent that can operate across different environments and perform various tasks, which could be useful for automating digital tasks in real-world applications. Practitioners may care about this work because it addresses the challenges of deploying multimodal GUI agents in practical settings.
This paper introduces a new AI model called T1, which uses reinforcement learning to perform long-term tasks such as coding and scientific discovery. Practitioners may care about this model because it can perform better than existing models on long-term tasks.
This paper provides a comprehensive survey of progress reward modeling in robotic learning, aiming to bridge the gap in the field by offering a unified framework for understanding progress rewards. Practitioners in robotics and AI can care about this paper because it helps them understand the different approaches to progress rewards and how to evaluate their effectiveness.
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 proposes a model called TLive-Omni that can understand and process multiple types of data (image, video, audio, text) from e-commerce live streaming. A practitioner might care about this model because it can be used to analyze and respond to live-commerce streams in a more accurate and efficient way.
This paper introduces FlowBalance, a method that helps a reasoning model improve itself by learning from its own experiences, while avoiding overconfidence and focusing on the best solutions. Practitioners might care about this approach because it can lead to more accurate and diverse model performance.
This paper addresses a problem in self-distillation, where a student model learns to mimic the behavior of a privileged teacher, but performs poorly at inference due to a "privilege illusion". The authors propose a new method, Dual-Anchored Policy Distillation, to resolve this issue and improve performance.
This paper introduces Intern-S2-Preview, a type of AI model designed to support scientific discovery by reasoning over different types of evidence, interacting with scientific tools, and sustaining progress over long periods. Practitioners might care because it could lead to more accurate scientific understanding and forecasting.
This paper presents a method for end-to-end driving systems to learn beyond human demonstrations, using a combination of perception and action models that can learn from diverse data and provide goal-conditioned supervision. Practitioners might care about this approach because it allows for more robust and flexible autonomous driving systems.
This paper develops a system to automatically optimize the design of harnesses for LLM agents, which can improve their reliability on long-horizon tasks. Practitioners might care about this because it could lead to more robust and performant agent systems.
This paper proposes a new approach to agentic visual reasoning, which helps large language models (LLMs) perform better on complex tasks by using tools more efficiently. Practitioners might care about this research because it aims to improve the performance of LLMs on challenging problems.
This paper proposes a reinforcement learning method to generate high-quality agent skills by progressively editing skills in a sequential process. Practitioners may care about this approach because it can improve the performance of agents on downstream tasks.
This paper introduces SPADE, a self-play framework that enables language agents to learn from adaptive, self-generated environments, allowing them to improve continuously without fixed goal distributions. Practitioners might care because SPADE can lead to more robust and open-ended AI models.
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.