The paper investigates how data should be scaled in agentic reinforcement learning, where data is now a structured design space of tasks, environments, and trajectories. By fitting per‑axis scaling laws across various large language models and RL algorithms, the authors find that environment diversity has the strongest impact, followed by task and trajectory diversity. Using these laws, they price synthesis methods, identify critical synthetic‑to‑real ratios, and formulate a compute‑optimal budget allocation that can improve efficiency by up to 1.7× compared to the common practice of scaling trajectories alone.
arXiv:2606. 18284v1 Announce Type: cross Abstract: The limiting resource for training agents via reinforcement learning (RL) is increasingly frontier task supply: valid, solvable tasks just difficult enough to train the current model.
arXiv:2610.29458v1 Announce Type: new Abstract: Agent harnesses have become a central component of modern agentic systems, shaping how models use tools, manage context, and coordinate long-horizon in...
arXiv:2507. 04136v2 Announce Type: replace Abstract: This survey offers a comprehensive foundation on the integration of RL with language models, highlighting prominent algorithms such as Proximal Policy Optimization (PPO), Q-Learning, and Actor-Critic methods.
MiMo‑V2.6 is an omni‑modal reinforcement learning framework that scales compute to advance large foundation models toward self‑improvement. It trains on a broad multimodal corpus before RL, then scales RL along three axes: larger batches and throughput, diverse environments across code, general, visual, and cyber domains, and enhanced grader compute for better reward signals. The system incorporates infrastructure such as a frozen MoE router, multi‑layer defense against reward hacking, unified trajectory representation, high‑concurrency rollout, and training‑inference consistency, and it is open‑source to support reproducibility and further research.
The paper introduces the Agentic Compositional Generalization hypothesis, suggesting that reinforcement learning (RL) primarily refines high‑level decision‑making behaviors that orchestrate pre‑trained low‑level skills, rather than teaching new domain‑specific skills from scratch. It proposes River, a training recipe that enhances reward quality by filtering low‑quality synthetic environments and adding process‑level behavior regularization. Using River, RL‑trained agents outperform other open‑source 8B models on four terminal‑agent benchmarks, achieving significant gains with fewer than 30% of the training environments.
arXiv:2606. 17454v1 Announce Type: new Abstract: AI agent performance is not just a modeling problem, it is fundamentally a systems problem.
arXiv:2610.28409v1 Announce Type: cross Abstract: Large language model agents are often given an inference budget and asked to maximise success within it. A multi-turn agent has to allocate this budg...
The paper introduces CoPE, a method that trains language agents by adding multi‑step action targets to the policy backbone. By supervising each action forecasted from preceding histories, CoPE enables acting and forecasting to co‑evolve without extra inference cost. Across five benchmarks, CoPE outperforms existing RL baselines, achieving up to twice the sample efficiency on ScienceWorld and improving overall performance on realistic applications.
arXiv:2608. 27260v1 Announce Type: new Abstract: LLM agents increasingly rely on generated interaction data to learn how to interact with external environments.
The paper presents a new scaling law for reward optimization in AI alignment, showing that performance scales as Θ(√min{log(M), K}), where M is the number of preference comparisons used to train a proxy reward model and K is the KL‑divergence budget relative to a reference policy. The authors derive this law using an information‑theoretic model, prove its tightness, and validate it with extensive experiments involving a 70B gold reward model and smaller proxy models (0.6B–4B). The empirical results demonstrate a strong fit (R² 97–99 %) across different model sizes, noise levels, and optimization methods, suggesting that reward optimization behaves like a simple selection task over IID Gaussian variables with noisy feedback.
The paper introduces Reinforcement Learning with Decomposed Subtasks (RLDS), a method that splits trajectory rewards into per‑subtask shares before policy updates, replacing the scalar advantage used in Group Relative Policy Optimization (GRPO). RLDS employs Subtask‑Decomposed Advantage Estimation (SDAE) to compute group‑relative advantages and distribute credit to tokens based on subtask importance, focusing on steps where a reflection marks a subtask as consequential. Experiments on four benchmarks—FrozenLake, HotpotQA, ScienceWorld, and DeepResearch—show that RLDS improves performance on high‑heterogeneity tasks (ScienceWorld and FrozenLake) and is more compute‑efficient than scalar GRPO for long rollouts.
arXiv:2609.36887v1 Announce Type: new Abstract: Recent efforts to scale tool-use post-training have largely centered on the synthesis of executable environments, which constitute only one component o...