Filtered to reinforcement learning · clear filter
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 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 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 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 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 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 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 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 introduces a new framework for building AI agents in Python, allowing developers to write agents that are both deterministic and model-agnostic. Practitioners might care because this framework could simplify the development of reliable AI agents for various applications.
This paper introduces Molt, a lightweight PyTorch framework for agentic reinforcement learning that aims to simplify the development process by reducing the overhead of algorithm modifications and framework changes. Practitioners might care about Molt because it can help them build and train reinforcement learning models more efficiently.
This paper addresses a problem in on-policy distillation where a student model can get stuck on a wrong path, and proposes a new method called Relay-OPD that helps the student model recover by briefly taking over at certain points to produce a new trajectory. Practitioners might care about this because it could lead to better performance and more efficient training in models like language generators or math solvers.
This paper proposes a method to improve reinforcement learning with verifiable rewards by using game solvers to provide turn-level credit to agents, allowing them to learn more effectively. Practitioners might care about this approach because it could lead to more robust and efficient AI decision-making.
This paper proposes a new method for reinforcement learning in large language models, called Group Entropy-Controlled Policy Optimization (GEPO), which helps balance exploration and exploitation by controlling entropy levels across different tasks. Practitioners might care about GEPO because it can lead to more balanced and task-specific exploration levels.
This paper proposes a method to combine reinforcement learning with verifiable rewards and on-policy distillation to improve performance on complex tasks, and shows that this method can lead to more stable training and better results.
This paper develops a new method for improving reasoning language models, called β-OPSD, which combines policy optimization and self-distillation to improve stability and performance. Practitioners might care about this method because it provides a more efficient and effective way to improve language model reasoning abilities.
This paper develops a framework, SpatialCLI, to help vision-language models (VLMs) better understand and use visual tools to make better decisions. By training VLMs to reason with spatial tools and then internalize those capabilities, SpatialCLI can improve the performance of VLMs in tasks that require visual reasoning.
This paper improves a type of reinforcement learning (RL) called rubric-based RL, which helps large language models (LLMs) perform well on open-ended tasks. A practitioner might care about this paper because it addresses a common problem in RL, where some criteria (or rules) are not explored properly, and it shows that its new method can improve performance on these tasks.
This paper proposes a new approach to memory-augmentation in large language model agents, allowing them to actively reconstruct and adapt past experiences to fit the current context, rather than simply replaying them. Practitioners might care because this approach can improve the robustness and intrinsic reasoning capabilities of agents in complex scenarios.
This paper investigates the use of the Muon optimizer in reinforcement learning (RL) post-training and finds that it can significantly improve the success rate of RL agents, especially when combined with other techniques like policy optimization and advantage estimation. Practitioners in RL may care about this research to explore new ways to improve the performance of their agents.
This paper addresses a problem in reinforcement learning (RL) where algorithms with PPO-Clip suffer from "exploration collapse", where they fail to explore new areas of the environment. The authors propose a new method called Riemannian Isometric Policy Optimization (RIPO) that helps to balance exploration and exploitation by using a more accurate geometric metric.
This paper proposes a method to improve on-policy distillation by reusing pre-collected teacher data, allowing for faster training without interacting with the environment. Practitioners may care about this technique because it enables scalable and efficient distillation of complex agent models.
This paper develops a new method to evaluate and verify image editing consistency across multiple references, addressing a challenge in reinforcement learning for multi-reference editing. Practitioners may care about this approach as it enables more accurate and reliable reinforcement learning for image editing tasks.
This paper proposes a new method for advantage shaping in reinforcement learning called Contrastive Policy Optimization (CPO), which uses contrastive disagreement between reference-guided and vanilla generation distributions to indicate correctness. Practitioners might care because it can improve the effectiveness of reinforcement learning methods in generating correct responses.
This paper proposes a new way for large language models to learn from feedback, allowing them to retain more detailed information about the quality of their responses and learn from it in a more nuanced way. Practitioners might care because this approach could lead to better performance on tasks where the model doesn't have a clear way to evaluate its own output.
This paper proposes a new method to improve reinforcement learning for large language models by using a predictive divergence mask, which helps stabilize off-policy updates. Practitioners caring about improving the stability and performance of LLM-based reinforcement learning might find this approach useful.
This paper uses reinforcement learning to optimize code for speed while maintaining correctness, and it finds that the approach can significantly improve performance, even when the reward function is noisy or sparse. Practitioners might care because optimizing code for speed can be crucial in many applications, and this research provides a promising method for achieving that goal.
This paper proposes a new method called Distilled Reinforcement Learning that improves large language model post-training by providing fine-grained guidance to transfer new knowledge from a teacher model to a student model. Practitioners might care because it outperforms standard reinforcement learning and on-policy distillation methods in terms of knowledge transfer and model performance.
This paper introduces Robostral Navigate, a vision-language model that enables robots to navigate using only a single monocular RGB camera, making it more scalable and cost-effective for deployment across various robotic platforms. Practitioners might care about this because it can simplify navigation tasks for robots in real-world environments.
This paper creates a new framework, OpenForgeRL, that allows researchers to train AI agents in complex environments using real harnesses, rather than relying on simplified inference systems. Practitioners might care because it enables more realistic testing and training of agents in real-world settings.
This paper develops a new type of artificial intelligence model that can simulate complex environments and make decisions in them, which could be useful for training robots and other agents to perform tasks in the real world. Practitioners might care because these models could help solve problems in areas like robotics, autonomous vehicles, and healthcare.
This paper develops a new optimization framework called Isospectral Optimization (ISO) that improves the performance of language models trained using reinforcement learning with verifiable rewards (RLVR) by reusing the base model's weight spectra while adapting the input and output singular frames. Practitioners might care about this because it could lead to faster and more accurate training of large language models.
This paper introduces H^2SD, a hybrid hindsight self-distillation framework for reinforcement learning with verifiable rewards, which combines the strengths of different methods to improve large language models' reasoning capabilities. Practitioners may care about this work because it addresses limitations of existing methods and shows promising results on challenging reasoning benchmarks.
This paper introduces a new environment called Trace, which allows vision-language models to reason across multiple domains and tasks, using a taxonomy-guided approach. Practitioners might care because this work could lead to more generalizable and transferable AI models.
This paper introduces Agon, a method for training reinforcement learning models to improve their reasoning abilities by competing with each other, rather than just optimizing for the final answer. Practitioners may care about using this approach to improve the quality of reasoning in AI models.