Large language model agents in social deduction games must reason and speak under hidden roles, shifting public beliefs, and social pressure. Existing methods mainly target final win rate through strategic prompting, belief modeling, or reinforcement learning, but they leave the utterance-level trade-off between belief correction and social risk largely implicit. In false-consensus intervention, useful private information can move the group toward the truth, while poorly timed or overly direct speech can trigger suspicion or rejection. We propose PACR-Pro, a risk-aware rhetorical decision framework for false-consensus intervention. PACR-Pro maps each speaking step to consensus-conflict encoding, structured candidate-action generation, Pareto-based selection, and risk-adaptive utterance realization. A dual-branch Epistemic-Social Critic scores candidate actions, and a state-conditioned 𝜆 policy adjusts the trade-off between epistemic value and social safety. We train these decision modules offline from historical games and evaluate PACR-Pro on One Night Ultimate Werewolf (ONUW) as the test platform. Experiments on standard ONUW games and the Wolf-Pack Induction Benchmark show that PACR-Pro achieves the best mixed-agent win rate, the highest attack win rate, the lowest attack vote rate, and the smallest attack-control win-rate gap among all baselines.

Large language models (LLMs) have significantly advanced automated code generation and debugging, facilitating powerful multi-agent coding frameworks. However, deploying these sophisticated models on resource-constrained edge devices remains challenging due to high computational demands, limited adaptability, and significant privacy risks associated with cloud-based processing. Motivated by these constraints, we propose Edge Code Cloak Coder (ECCC), a novel edge-cloud hybrid framework integrating lightweight quantized LLM with robust AST-based anonymization and edge-side privacy validation. ECCC enables high-performance, privacy-preserving LLM capabilities on consumer GPUs, anonymizing user code before securely delegating abstracted tasks to cloud LLMs. Experimental evaluations demonstrate that ECCC achieves competitive correctness (within 4–5pp of the GPT-4-based frameworks) and a perfect privacy score of 10/10, effectively balancing functionality and security for sensitive and proprietary code applications.

“神经网络模型的快速发展使得多文档摘要可以获得人类可读的流畅的摘要,对大规模的数据进行预训练可以更好的从自然语言文本中捕捉更丰富的语义信息,并更好的作用于下游任务。目前很多的多文档摘要的工作也应用了预训练模型(如BERT)并取得了一定的效果,但是这些预训练模型不能更好的从文本中捕获事实性知识,没有考虑到多文档文本的结构化的实体-关系信息,本文提出了基于实体信息增强和多粒度融合的多文档摘要模型MGNIE,将实体关系信息融入预训练模型ERNIE中,增强知识事实以获得多层语义信息,解决摘要生成的事实一致性问题。进而从多种粒度进行多文档层次结构的融合建模,以词信息、实体信息以及句子信息捕捉长文本信息摘要生成所需的关键信息点。本文设计的模型,在国际标准评测数据集MultiNews上对比强基线模型效果和竞争力获得较大提升。”