In heterogeneous multi-task decision-making, tasks not only exhibit diverse observation and action spaces but also vary substantially in their underlying complexities. While conventional multi-task world models like UniZero excel in single-task settings, we find that when handling a broad and diverse suite of tasks, gradient conflicts and the loss of model plasticity often constrain their sample efficiency. In this work, we address these challenges from two complementary perspectives: the single learning iteration and the overall learning process. First, to mitigate the gradient conflicts, we systematically investigate key architectural designs for extending UniZero. Our investigation identifies a Mixture-of-Experts (MoE) architecture as the most effective approach. We demonstrate, both theoretically and empirically, that this architecture alleviates gradient conflicts by routing task-specific representations to specialized sub-networks. This finding leads to our proposed model, \textit{ScaleZero}. Second, to dynamically allocate model capacity throughout the learning process, we introduce an online Dynamic Parameter Scaling (DPS) strategy. This strategy progressively integrates LoRA adapters in response to task-specific progress, enabling adaptive knowledge retention and parameter expansion. Evaluations on a diverse set of standard benchmarks (Atari, DMC, Jericho) demonstrate that ScaleZero, utilizing solely online reinforcement learning with one model, performs on par with specialized single-task agents. With the DPS strategy, it remains competitive while using just 71.5% of the environment interactions. These findings underscore the potential of ScaleZero for effective multi-task planning. Our code is available at \textcolor{magenta}{https://github.com/opendilab/LightZero}.

For RL algorithms, appropriate entropy control is crucial to their effectiveness. To control the policy entropy, a commonly used method is entropy regularization, which is adopted in various popular RL algorithms including PPO, SAC and A3C. Although entropy regularization proves effective in robotic and games RL conventionally, studies found that it gives weak to no gains in LLM-RL training. In this work, we study the issues of entropy bonus in LLM-RL setting. Specifically, we first argue that the conventional entropy regularization suffers from the LLM's extremely large response space and the sparsity of the optimal outputs. As a remedy, we propose AEnt, an entropy control method that utilizes a new clamped entropy bonus with an automatically adjusted coefficient. The clamped entropy is evaluated with the re-normalized policy defined on certain smaller token space, which encourages exploration within a more compact response set. In addition, the algorithm automatically adjusts entropy coefficient according to the clamped entropy value, effectively controlling the entropy-induced bias while leveraging the entropy's benefits. AEnt is tested in math-reasoning tasks under different base models and datasets, and it is observed that AEnt outperforms the baselines consistently across multiple benchmarks.

Diffusion language models, as a promising alternative to traditional autoregressive (AR) models, enable faster generation and richer conditioning on bidirectional context. However, they suffer from a key discrepancy between training and inference: during inference, MDLMs progressively reveal the structure of the generated sequence by producing fewer and fewer masked tokens, whereas this structure is ignored in training as tokens are masked at random. Although this discrepancy between training and inference can lead to suboptimal performance, it has been largely overlooked by previous works, leaving closing this gap between the two stages an open problem. To address this, we frame the problem of learning effective denoising trajectories as a sequential decision-making problem and use the resulting framework to apply reinforcement learning. We propose a novel Masked Diffusion Policy Optimization (MDPO) to exploit the Markov property diffusion possesses and explicitly train the model under the same progressive refining schedule used at inference. MDPO matches the performance of the previous state-of-the-art (SOTA) method with 60x fewer gradient updates, while achieving average improvements of 9.6% on MATH500 and 54.2% on Countdown over SOTA when trained within the same number of weight updates. Additionally, we improve the remasking strategy of MDLMs as a plug-in inference replacement to overcome the limitation that the model cannot refine tokens flexibly. This training-free method, termed Running Confidence Remasking (RCR), consistently enhances performance and provides further improvements when used with MDPO. Our findings establish great potential for investigating the discrepancy between pre-training and inference of MDLMs. Code: https://github.com/autonomousvision/mdpo. Project Page: https://cli212.github.io/MDPO/.

While LLMs exhibit impressive fluency and factual recall, they struggle with robust causal reasoning, often relying on spurious correlations and brittle patterns. Similarly, traditional Reinforcement Learning agents also lack causal understanding, optimizing for rewards without modeling why actions lead to outcomes. We introduce Causal Reflection, a framework that explicitly models causality as a dynamic function over state, action, time, and perturbation, enabling agents to reason about delayed and nonlinear effects. Additionally, we define a formal Reflect mechanism that identifies mismatches between predicted and observed outcomes and generates causal hypotheses to revise the agent's internal model. In this architecture, LLMs serve not as black-box reasoners, but as structured inference engines translating formal causal outputs into natural language explanations and counterfactuals. Our framework lays the theoretical groundwork for Causal Reflective agents that can adapt, self-correct, and communicate causal understanding in evolving environments.

We propose VRAIL (Vectorized Reward-based Attribution for Interpretable Learning), a bi-level framework for value-based reinforcement learning (RL) that learns interpretable weight representations from state features. VRAIL consists of two stages: a deep learning (DL) stage that fits an estimated value function using state features, and an RL stage that uses this to shape learning via potential-based reward transformations. The estimator is modeled in either linear or quadratic form, allowing attribution of importance to individual features and their interactions. Empirical results on the Taxi-v3 environment demonstrate that VRAIL improves training stability and convergence compared to standard DQN, without requiring environment modifications. Further analysis shows that VRAIL uncovers semantically meaningful subgoals, such as passenger possession, highlighting its ability to produce human-interpretable behavior. Our findings suggest that VRAIL serves as a general, model-agnostic framework for reward shaping that enhances both learning and interpretability.

Agents trained with reinforcement learning often develop brittle policies that fail when dynamics shift, a problem amplified by static benchmarks. AbideGym, a dynamic MiniGrid wrapper, introduces agent-aware perturbations and scalable complexity to enforce intra-episode adaptation. By exposing weaknesses in static policies and promoting resilience, AbideGym provides a modular, reproducible evaluation framework for advancing research in curriculum learning, continual learning, and robust generalization.

Protecting privacy in social graphs requires preventing sensitive information, such as community affiliations, from being inferred by graph analysis, without substantially altering the graph topology. We address this through the problem of \emph{community membership hiding} (CMH), which seeks edge modifications that cause a target node to exit its original community, regardless of the detection algorithm employed. Prior work has focused on non-overlapping community detection, where trivial strategies often suffice, but real-world graphs are better modeled by overlapping communities, where such strategies fail. To the best of our knowledge, we are the first to formalize and address CMH in this setting. In this work, we propose a deep reinforcement learning (DRL) approach that learns effective modification policies, including the use of proxy nodes, while preserving graph structure. Experiments on real-world datasets show that our method significantly outperforms existing baselines in both effectiveness and efficiency, offering a principled tool for privacy-preserving graph modification with overlapping communities.

Large Language Models (LLMs) have been used to make decisions in complex scenarios, where they need models to think deeply, reason logically, and decide wisely. Many existing studies focus solely on multi-round conversations in social tasks or simulated environments, neglecting the various types of decisions and their interdependence. Current reinforcement learning methods struggle to consider the strategies of others during training. To address these issues, we first define a strategic decision-making problem that includes two types of decisions and their temporal dependencies. Furthermore, we propose **T**heory **o**f **M**ind **P**olicy **O**ptimization **(ToMPO)** algorithm to optimize the perception of other individual strategies and the game situation trends. Compared to the Group Relative Policy Optimization (GRPO) algorithm, ToMPO enhances the LLM's strategic decision-making mainly by: 1) generating rollouts based on reasoning the strategies of other individuals, 2) estimating advantages at both the graph-level and sample-level, and 3) balancing global and partial rewards. The ToMPO algorithm outperforms the GRPO method by 35% in terms of model output compliance and cooperative outcomes. Additionally, when compared to models with parameter sizes 100 times larger, it shows an 18% improvement. This demonstrates the effectiveness of the ToMPO algorithm in enhancing the model's strategic decision-making capabilities.

Online reinforcement learning in complex tasks is time-consuming, as massive interaction steps are needed to learn the optimal Q-function.Vision-language action (VLA) policies represent a promising direction for solving diverse tasks; however, their performance on low-level control remains limited, and effective deployment often requires task-specific expert demonstrations for fine-tuning. In this paper, we propose \textbf{VARL} (\textbf{V}LM as \textbf{A}ction advisor for online \textbf{R}einforcement \textbf{L}earning), a framework that leverages the domain knowledge of vision-language models (VLMs) to provide action suggestions for reinforcement learning agents. Unlike previous methods, VARL provides action suggestions rather than designing heuristic rewards, thereby guaranteeing unchanged optimality and convergence. The suggested actions increase sample diversity and ultimately improve sample efficiency, especially in sparse-reward tasks. To validate the effectiveness of VARL, we evaluate it across diverse environments and agent settings. Results show that VARL greatly improves sample efficiency without introducing significant computational overhead. These advantages make VARL a general framework for online reinforcement learning and make it feasible to directly apply reinforcement learning from scratch in real-world environments.

Abstract-- This paper addresses a drone ball-balancing task, in which a drone stabilizes a ball atop a movable beam through cable-based interaction. We propose a hierarchical control framework that decouples high-level balancing policy from low-level drone control, and train a reinforcement learning (RL) policy to handle the high-level decision-making. Simulation results show that the RL policy achieves superior performance compared to carefully tuned PID controllers within the same hierarchical structure. Through systematic comparative analysis, we demonstrate that RL's advantage stems not from improved parameter tuning or inherent nonlinear mapping capabilities, but from its ability to effectively utilize richer state observations. These findings underscore the critical role of comprehensive state representation in learning-based systems and suggest that enhanced sensing could be instrumental in improving controller performance.