Caught in the Story: Narrative Captivity in Multi-turn LLMs Conversation
研究解决了多轮对话中大型语言模型易受单方面叙述影响的问题,通过构建包含5078个情景的基准测试,发现17个模型普遍存在此问题,并探索了部分缓解策略。
研究解决了多轮对话中大型语言模型易受单方面叙述影响的问题,通过构建包含5078个情景的基准测试,发现17个模型普遍存在此问题,并探索了部分缓解策略。
该研究提出SUP-MIMIC框架,通过设计DDT和DCT任务评估大型语言模型在处理临床诊断中矛盾证据时的推理能力,揭示了模型依赖统计捷径而非因果推理的问题。
为解决MPCA在重尾分布和污染下的不稳定性,提出基于空间符号的多线性主成分分析(SMPCA),通过空间中位数中心化、空间符号归一化及交替特征分解来稳健降维。
This work addresses the limitations of existing edge service orchestration approaches, which overlook the reorderability of local sequences within service function chain directed acyclic graphs (SFC-DAGs) and fail to jointly optimize request scheduling and resource allocation. To bridge this gap, the authors propose RQ-SAFE, a novel framework that, for the first time, couples the flexibility of SFC-DAG local ordering with queue-aware resource scheduling. RQ-SAFE dynamically evaluates the resource impact of feasible local orderings in an online manner and leverages real-time queue states to guide virtual network function (VNF) instance selection and path construction. Furthermore, it incorporates a learning-assisted reordering mechanism to balance quality of service (QoS) and system load. Experimental results demonstrate that, compared to the GNN-DAG-Score baseline, RQ-SAFE reduces CPU load imbalance by 6.1%, lowers peak utilization by 2.3%, and improves QoS by 4.53 percentage points through joint optimization.
Existing LLM-based trading agents predominantly rely on single-step prediction and lack explicit risk management mechanisms, leading to suboptimal performance under market volatility. To address this, we propose RISK-LLM—a novel framework featuring: (i) hierarchical market analysis to model multi-granularity dynamics; (ii) a dual-decision agent architecture that decouples signal generation from risk control; and (iii) multi-horizon reinforcement learning with risk-aware rewards, jointly optimizing returns and downside risk constraints (e.g., conditional value-at-risk). This work constitutes the first systematic integration of risk sensitivity into the LLM-driven trading paradigm. Extensive experiments across diverse markets—including A-shares and U.S. equities—demonstrate that RISK-LLM significantly improves the Sharpe ratio and enhances maximum drawdown control, outperforming state-of-the-art methods in both profitability and trading stability.
研究解决了多轮对话中大型语言模型易受单方面叙述影响的问题,通过构建包含5078个情景的基准测试,发现17个模型普遍存在此问题,并探索了部分缓解策略。
该研究提出SUP-MIMIC框架,通过设计DDT和DCT任务评估大型语言模型在处理临床诊断中矛盾证据时的推理能力,揭示了模型依赖统计捷径而非因果推理的问题。
为解决MPCA在重尾分布和污染下的不稳定性,提出基于空间符号的多线性主成分分析(SMPCA),通过空间中位数中心化、空间符号归一化及交替特征分解来稳健降维。
This work addresses the limitations of existing edge service orchestration approaches, which overlook the reorderability of local sequences within service function chain directed acyclic graphs (SFC-DAGs) and fail to jointly optimize request scheduling and resource allocation. To bridge this gap, the authors propose RQ-SAFE, a novel framework that, for the first time, couples the flexibility of SFC-DAG local ordering with queue-aware resource scheduling. RQ-SAFE dynamically evaluates the resource impact of feasible local orderings in an online manner and leverages real-time queue states to guide virtual network function (VNF) instance selection and path construction. Furthermore, it incorporates a learning-assisted reordering mechanism to balance quality of service (QoS) and system load. Experimental results demonstrate that, compared to the GNN-DAG-Score baseline, RQ-SAFE reduces CPU load imbalance by 6.1%, lowers peak utilization by 2.3%, and improves QoS by 4.53 percentage points through joint optimization.
Existing LLM-based trading agents predominantly rely on single-step prediction and lack explicit risk management mechanisms, leading to suboptimal performance under market volatility. To address this, we propose RISK-LLM—a novel framework featuring: (i) hierarchical market analysis to model multi-granularity dynamics; (ii) a dual-decision agent architecture that decouples signal generation from risk control; and (iii) multi-horizon reinforcement learning with risk-aware rewards, jointly optimizing returns and downside risk constraints (e.g., conditional value-at-risk). This work constitutes the first systematic integration of risk sensitivity into the LLM-driven trading paradigm. Extensive experiments across diverse markets—including A-shares and U.S. equities—demonstrate that RISK-LLM significantly improves the Sharpe ratio and enhances maximum drawdown control, outperforming state-of-the-art methods in both profitability and trading stability.