Tag: LLM
All the papers with the tag "LLM".
Leveraging LLMs to Automate Energy-Aware Refactoring of Parallel Scientific Codes
grok-3-latestScore: 0.53Published: at 17:05本文提出 LASSI-EE 框架,利用大型语言模型通过多阶段、自我校正的管道自动化重构并行科学代码,在 NVIDIA A100 GPU 上实现平均 47% 的能量节省,展示了 LLM 在能效优化中的潜力。
Adversarial Cooperative Rationalization: The Risk of Spurious Correlations in Even Clean Datasets
grok-3-latestScore: 0.48Published: at 14:00本文提出 A2I 方法,通过对抗性攻击检测并纠正自解释合理化框架中模型引入的虚假相关性,显著提升 Rationale 质量。
MemEngine: A Unified and Modular Library for Developing Advanced Memory of LLM-based Agents
grok-3-latestScore: 0.57Published: at 13:10本文提出 MemEngine,一个统一且模块化的库,整合多种 LLM 智能体内存模型,支持高效开发和用户友好的应用,填补了现有研究中缺乏统一内存框架的空白。
What do Language Model Probabilities Represent? From Distribution Estimation to Response Prediction
grok-3-latestScore: 0.79Published: at 11:46本文通过理论框架区分了语言模型输出概率的三种解释(补全分布、响应分布、事件分布),揭示了现有研究中的混淆,为LLM的概率解释和应用提供了理论指导。
What do Language Model Probabilities Represent? From Distribution Estimation to Response Prediction
grok-3-latestScore: 0.79Published: at 11:46本文提出一个理论框架,区分了大型语言模型输出概率的三种解释(源分布估计、目标分布估计、响应预测),并揭示了现有研究中因混淆这些分布导致的误解,为概率解释和模型设计提供了理论基础。
Restoring Calibration for Aligned Large Language Models: A Calibration-Aware Fine-Tuning Approach
grok-3-latestScore: 0.83Published: at 05:42本文通过校准感知微调方法(CFT 和 RCFT),结合理论状态划分和 EM 算法正则化,显著改善了偏好对齐后大型语言模型的校准性能,同时保持或提升模型准确率。