Despite the impressive capabilities of multimodal large language models (MLLMs) in vision-language tasks, they are prone to hallucinations in real-world scenarios. This paper investigates the hallucination phenomenon in MLLMs from the perspective of modality conflict. Unlike existing works focusing on the conflicts between model responses and inputs, we study the inherent conflicts in inputs from different modalities that place MLLMs in a dilemma and directly lead to hallucinations. We formally define the modality conflict and construct a dataset named Multimodal Modality Conflict (MMMC) to simulate this phenomenon in vision-language tasks. Three methods based on prompt engineering, supervised fine-tuning, and reinforcement learning are proposed to alleviate the hallucination caused by modality conflict. Extensive experiments are conducted on the MMMC dataset to analyze the merits and demerits of these methods. Our results show that the reinforcement learning method achieves the best performance in mitigating the hallucination under modality conflict, while the supervised fine-tuning method shows promising and stable performance. Our work sheds light on the unnoticed modality conflict that leads to hallucinations and provides more insights into the robustness of MLLMs.

We introduce a real-time strategy game environment based on Generals.io, a game with thousands of weekly active players. Our environment is fully compatible with Gymnasium and PettingZoo and is capable of running thousands of frames per second on commodity hardware. We also present a reference agent, trained with supervised pre-training and self-play, which reached the top 0.003% of the 1v1 human leaderboard after only 36 hours on a single H100 GPU. To accelerate learning, we incorporate potential-based reward shaping and memory features. Our contributions of a modular RTS benchmark and a competitive baseline agent provide an accessible yet challenging platform for advancing multi-agent reinforcement learning research. The documented code, together with examples and tutorials, is available at https://github.com/strakam/generals-bots.

The integration of emerging uncrewed aerial vehicles (UAVs) with artificial intelligence (AI) and ground-embedded robots (GERs) has transformed emergency rescue operations in unknown environments. However, the high computational demands often exceed a single UAV's capacity, making it difficult to continuously provide stable high-level services. To address this, this paper proposes a cooperation framework involving UAVs, GERs, and airships. The framework enables resource pooling through UAV-to-GER (U2G) and UAV-to-airship (U2A) links, offering computing services for offloaded tasks. Specifically, we formulate the multi-objective problem of task assignment and exploration as a dynamic long-term optimization problem aiming to minimize task completion time and energy use while ensuring stability. Using Lyapunov optimization, we transform it into a per-slot deterministic problem and propose HG-MADDPG, which combines the Hungarian algorithm with a GDM-based multi-agent deep deterministic policy gradient. Simulations demonstrate significant improvements in offloading efficiency, latency, and system stability over baselines.

This article presents Platform Adaptive Locomotion (PAL), a unified control method for quadrupedal robots with different morphologies and dynamics. We leverage deep reinforcement learning to train a single locomotion policy on procedurally generated robots. The policy maps proprioceptive robot state information and base velocity commands into desired joint actuation targets, which are conditioned using a latent embedding of the temporally local system dynamics. We explore two conditioning strategies - one using a GRU-based dynamics encoder and another using a morphology-based property estimator - and show that morphology-aware conditioning outperforms temporal dynamics encoding regarding velocity task tracking for our hardware test on ANYmal C. Our results demonstrate that both approaches achieve robust zero-shot transfer across multiple unseen simulated quadrupeds. Furthermore, we demonstrate the need for careful robot reference modelling during training: exposing the policy to a diverse set of robot morphologies and dynamics leads to improved generalization, reducing the velocity tracking error by up to 30% compared to the baseline method. Despite PAL not surpassing the best-performing reference-free controller in all cases, our analysis uncovers critical design choices and informs improvements to the state of the art.

Reinforcement learning (RL) for large language models is an energy-intensive endeavor: training can be unstable, and the policy may gradually drift away from its pretrained weights. We present \emph{RLEP}\, -- \,Reinforcement Learning with Experience rePlay\, -- \,a two-phase framework that first collects verified trajectories and then replays them during subsequent training. At every update step, the policy is optimized on mini-batches that blend newly generated rollouts with these replayed successes. By replaying high-quality examples, RLEP steers the model away from fruitless exploration, focuses learning on promising reasoning paths, and delivers both faster convergence and stronger final performance. On the Qwen2.5-Math-7B base model, RLEP reaches baseline peak accuracy with substantially fewer updates and ultimately surpasses it, improving accuracy on AIME-2024 from 38.2% to 39.9%, on AIME-2025 from 19.8% to 22.3%, and on AMC-2023 from 77.0% to 82.2%. Our code, datasets, and checkpoints are publicly available at https://github.com/Kwai-Klear/RLEP to facilitate reproducibility and further research.

In this paper, deceptive signal-assisted private split learning is investigated. In our model, several edge devices jointly perform collaborative training, and some eavesdroppers aim to collect the model and data information from devices. To prevent the eavesdroppers from collecting model and data information, a subset of devices can transmit deceptive signals. Therefore, it is necessary to determine the subset of devices used for deceptive signal transmission, the subset of model training devices, and the models assigned to each model training device. This problem is formulated as an optimization problem whose goal is to minimize the information leaked to eavesdroppers while meeting the model training energy consumption and delay constraints. To solve this problem, we propose a soft actor-critic deep reinforcement learning framework with intrinsic curiosity module and cross-attention (ICM-CA) that enables a centralized agent to determine the model training devices, the deceptive signal transmission devices, the transmit power, and sub-models assigned to each model training device without knowing the position and monitoring probability of eavesdroppers. The proposed method uses an ICM module to encourage the server to explore novel actions and states and a CA module to determine the importance of each historical state-action pair thus improving training efficiency. Simulation results demonstrate that the proposed method improves the convergence rate by up to 3x and reduces the information leaked to eavesdroppers by up to 13% compared to the traditional SAC algorithm.

In this paper, a secure and communication-efficient clustered federated learning (CFL) design is proposed. In our model, several base stations (BSs) with heterogeneous task-handling capabilities and multiple users with non-independent and identically distributed (non-IID) data jointly perform CFL training incorporating differential privacy (DP) techniques. Since each BS can process only a subset of the learning tasks and has limited wireless resource blocks (RBs) to allocate to users for federated learning (FL) model parameter transmission, it is necessary to jointly optimize RB allocation and user scheduling for CFL performance optimization. Meanwhile, our considered CFL method requires devices to use their limited data and FL model information to determine their task identities, which may introduce additional communication overhead. We formulate an optimization problem whose goal is to minimize the training loss of all learning tasks while considering device clustering, RB allocation, DP noise, and FL model transmission delay. To solve the problem, we propose a novel dynamic penalty function assisted value decomposed multi-agent reinforcement learning (DPVD-MARL) algorithm that enables distributed BSs to independently determine their connected users, RBs, and DP noise of the connected users but jointly minimize the training loss of all learning tasks across all BSs. Different from the existing MARL methods that assign a large penalty for invalid actions, we propose a novel penalty assignment scheme that assigns penalty depending on the number of devices that cannot meet communication constraints (e.g., delay), which can guide the MARL scheme to quickly find valid actions, thus improving the convergence speed. Simulation results show that the DPVD-MARL can improve the convergence rate by up to 20% and the ultimate accumulated rewards by 15% compared to independent Q-learning.

Efficient exploration is a well known problem in deep reinforcement learning and this problem is exacerbated in multi-agent reinforcement learning due the intrinsic complexities of such algorithms. There are several approaches to efficiently explore an environment to learn to solve tasks by multi-agent operating in that environment, of which, the idea of expert exploration is investigated in this work. More specifically, this work investigates the application of large-language models as expert planners for efficient exploration in planning based tasks for multiple agents.

The recent focus and release of pre-trained models have been a key components to several advancements in many fields (e.g. Natural Language Processing and Computer Vision), as a matter of fact, pre-trained models learn disparate latent embeddings sharing insightful representations. On the other hand, Reinforcement Learning (RL) focuses on maximizing the cumulative reward obtained via agent's interaction with the environment. RL agents do not have any prior knowledge about the world, and they either learn from scratch an end-to-end mapping between the observation and action spaces or, in more recent works, are paired with monolithic and computationally expensive Foundational Models. How to effectively combine and leverage the hidden information of different pre-trained models simultaneously in RL is still an open and understudied question. In this work, we propose Weight Sharing Attention (WSA), a new architecture to combine embeddings of multiple pre-trained models to shape an enriched state representation, balancing the tradeoff between efficiency and performance. We run an extensive comparison between several combination modes showing that WSA obtains comparable performance on multiple Atari games compared to end-to-end models. Furthermore, we study the generalization capabilities of this approach and analyze how scaling the number of models influences agents' performance during and after training.

Off-Policy Evaluation (OPE) aims to estimate the value of a target policy using offline data collected from potentially different policies. In real-world applications, however, logged data often suffers from missingness. While OPE has been extensively studied in the literature, a theoretical understanding of how missing data affects OPE results remains unclear. In this paper, we investigate OPE in the presence of monotone missingness and theoretically demonstrate that the value estimates remain unbiased under ignorable missingness but can be biased under nonignorable (informative) missingness. To retain the consistency of value estimation, we propose an inverse probability weighted value estimator and conduct statistical inference to quantify the uncertainty of the estimates. Through a series of numerical experiments, we empirically demonstrate that our proposed estimator yields a more reliable value inference under missing data.