Automatic medical report generation can greatly reduce the workload of doctors, but it is often unreliable for real-world deployment. Current methods can write formally fluent sentences but may be factually flawed, introducing serious medical errors known as clinical hallucinations, which make them untrustworthy for diagnosis. To bridge this gap, we introduce HiMed-RL, a Hierarchical Medical Reward Learning Framework designed to explicitly prioritize clinical quality. HiMed-RL moves beyond simple text matching by deconstructing reward learning into three synergistic levels: it first ensures linguistic fluency at the token-level, then enforces factual grounding at the concept-level by aligning key medical terms with expert knowledge, and finally assesses high-level diagnostic consistency at the semantic-level using a specialized LLM verifier. This hierarchical reward is implemented via a Human-inspired Dynamic Reward Adjustment, a strategy which first teaches the model to learn basic facts before progressing to more complex diagnostic reasoning. Experimentally, HiMed-3B achieves state-of-the-art performance on both in-domain and out-of-domain benchmarks, particularly on the latter, with an improvement of 10.8% over the second-best baseline. Our work provides a robust paradigm for generating reports that not only improve fluency but clinical fine-grained quality.

Abstract Despite extensive efforts to reveal temporal patterns of citation dynamics for individual papers, the spatial–temporal patterns, which reflect a coupled process involving the accumulation of citations over time and their concentration or dispersion in the high‐dimensional knowledge space, remain largely unexplored. Drawing on millions of papers in seven large‐scale disciplines, we apply document representation algorithms capturing semantic associations to construct citation diffusion trajectories in the embedded disciplinary knowledge landscape. We find a tendency toward semantically localized citation mobility patterns, characterized by Gaussian‐distributed citation scopes and exponentially distributed citation distances. Further analysis investigates the role of early citation diffusion characteristics in shaping citation accumulation, revealing a positive correlation between initial citation diffusion breadth and citation growth, which significantly amplifies long‐term impact among papers with comparable early citation counts. We identify distinct spatial–temporal citation diffusion patterns for novel and disruptive papers. While highly cited papers attract significant attention both in the early stages and over time, their citation diffusion typically follows conventional patterns, with most citations spanning shorter knowledge distances. Novel papers and disruptive papers, despite garnering fewer and delayed citations, tend to attract longer‐distance citations early on and generate more extensive and distant impacts within the knowledge space.

The prisoner’s dilemma highlights the conflict between individual gains and collective well-being, emphasizing the need for superior strategies. This article explores strategy learning through reinforcement learning, avoiding the time-intensive offline methods like genetic algorithms. Given the significance of both player behaviors and states in creating a lifelike gaming environment, two primary strategies are introduced: Q-learning (QLPlayer), focusing on game behaviors, and reinforcement learning with a finite state machine (RLFSM), integrating behaviors and states. Furthermore, two greedy strategies are developed: payoff-greedy (PG) based on historical payoffs and epsilon-greedy (EG) based on fixed probability. Following that, QLPlayer_PG, QLPlayer_EG, RLFSM_PG, and RLFSM_EG strategies are designed to choose the action with the highest expected payoffs based on different greedy strategies, and continuously update the expected payoffs in different states through online learning. Our experiments demonstrate that QLPlayer_PG and RLFSM_PG significantly outperform classical strategies in achieving higher payoffs and cooperation, especially in noisy environments, by dynamically adapting to learned opponent behaviors. Notably, QLPlayer_PG and RLFSM_PG exhibit superior performance compared to QLPlayer_EG and RLFSM_EG, indicating that the reinforcement learning strategy combining payoff-greedy can adjust its behavior over time according to the payoff information of both sides of the game, ensuring adaptability in the game. The results emphasize the importance of incorporating behaviors, states, and payoffs into strategy learning. This helps reduce exploitation, promote cooperation, and improve understanding of chosen actions.

Inspired by the ideas of the ldquoglobal brainrdquo and Swarm Intelligence, a concept of ldquocommunity intelligencerdquo is suggested in the present paper, reflecting that some ldquointelligentrdquo features may emerge in a Web-mediated community from interactions and knowledge-transmissions between community members. A conceptual model of community intelligence is developed from two views. From the structural view, the community intelligent system is modeled as a knowledge supernetwork that is comprised of triple interwoven networks of the media network (the Web), the human network, and the knowledge networks. Furthermore, based on a dyad of knowledge in two forms of ldquoknowingrdquo and ldquoknowarerdquo, the dynamic view describes the basic mechanics of the formation and evolution of ldquocommunity intelligencerdquo. Two categories of relevant research issues are shortly discussed on the basis of the proposed conceptual model.