Rui Liu

Rui Liu, Dian Yu, Tong Zheng et al.

Reinforcement learning with verifiable rewards (RLVR) improves reasoning in large language models (LLMs) but struggles with exploration, an issue that still persists for multimodal LLMs (MLLMs). Current methods treat the visual input as a fixed, deterministic condition, overlooking a critical source of ambiguity and struggling to build policies robust to plausible visual variations. We introduce $\textbf{VOGUE (Visual Uncertainty Guided Exploration)}$, a novel method that shifts exploration from the output (text) to the input (visual) space. By treating the image as a stochastic context, VOGUE quantifies the policy's sensitivity to visual perturbations using the symmetric KL divergence between a "raw" and "noisy" branch, creating a direct signal for uncertainty-aware exploration. This signal shapes the learning objective via an uncertainty-proportional bonus, which, combined with a token-entropy bonus and an annealed sampling schedule, effectively balances exploration and exploitation. Implemented within GRPO on two model scales (Qwen2.5-VL-3B/7B), VOGUE boosts pass@1 accuracy by an average of 2.6% on three visual math benchmarks and 3.7% on three general-domain reasoning benchmarks, while simultaneously increasing pass@4 performance and mitigating the exploration decay commonly observed in RL fine-tuning. Our work shows that grounding exploration in the inherent uncertainty of visual inputs is an effective strategy for improving multimodal reasoning.

Rui Liu, Yu Shen, Peng Gao et al.

Multi-modal learning has become a crucial technique for improving the performance of machine learning applications across domains such as autonomous driving, robotics, and perception systems. However, in certain scenarios, particularly in resource-constrained environments, some modalities available during training may be absent during inference. While existing frameworks effectively utilize multiple data sources during training and enable inference with reduced modalities, they are primarily designed for single-agent settings. This poses a critical limitation in dynamic environments such as connected autonomous vehicles (CAV), where incomplete data coverage can lead to decision-making blind spots. Conversely, some works explore multi-agent collaboration but without addressing missing modality at test time. To overcome these limitations, we propose Collaborative Auxiliary Modality Learning (CAML), a novel multi-modal multi-agent framework that enables agents to collaborate and share multi-modal data during training, while allowing inference with reduced modalities during testing. Experimental results in collaborative decision-making for CAV in accident-prone scenarios demonstrate that CAML achieves up to a ${\bf 58.1}\%$ improvement in accident detection. Additionally, we validate CAML on real-world aerial-ground robot data for collaborative semantic segmentation, achieving up to a ${\bf 10.6}\%$ improvement in mIoU.

Ruijiao Luo, Chao Huang, Yuntao Peng et al.

In human-robot collaboration (HRC), human trust in the robot is the human expectation that a robot executes tasks with desired performance. A higher-level trust increases the willingness of a human operator to assign tasks, share plans, and reduce the interruption during robot executions, thereby facilitating human-robot integration both physically and mentally. However, due to real-world disturbances, robots inevitably make mistakes, decreasing human trust and further influencing collaboration. Trust is fragile and trust loss is triggered easily when robots show incapability of task executions, making the trust maintenance challenging. To maintain human trust, in this research, a trust repair framework is developed based on a human-to-robot attention transfer (H2R-AT) model and a user trust study. The rationale of this framework is that a prompt mistake correction restores human trust. With H2R-AT, a robot localizes human verbal concerns and makes prompt mistake corrections to avoid task failures in an early stage and to finally improve human trust. User trust study measures trust status before and after the behavior corrections to quantify the trust loss. Robot experiments were designed to cover four typical mistakes, wrong action, wrong region, wrong pose, and wrong spatial relation, validated the accuracy of H2R-AT in robot behavior corrections; a user trust study with $252$ participants was conducted, and the changes in trust levels before and after corrections were evaluated. The effectiveness of the human trust repairing was evaluated by the mistake correction accuracy and the trust improvement.