Skill Issue: Lessons from Optimizing Repository SKILLs for Coding Agents
研究通过优化SKILL文件来提高编码代理性能,使用合并的拉取请求作为更难任务,并基于代理表现评分,发现GEPA方法有效提升了文档质量。
研究通过优化SKILL文件来提高编码代理性能,使用合并的拉取请求作为更难任务,并基于代理表现评分,发现GEPA方法有效提升了文档质量。
该文介绍了一种使用大型语言模型辅助开发静态验证软件的工具Eiffel-tools,通过与语言服务器协议结合,并利用形式化验证器提高代码生成和修正的准确性。
本文针对异构敏捷地球观测卫星调度问题,提出了一种结合强化学习引导的进化策略优化框架,通过解码器和基于群体的搜索方法实现高效优化。
This study addresses the limitations of holistic scoring in metaphor explanation evaluation, which often overlooks quality structure and human disagreement. We propose a cognition-driven, six-dimensional assessment framework to capture these nuances. Through large-scale annotation and clustering analysis, we reveal the multidimensionality of explanation quality and systematic patterns of disagreement, validating that an automated evaluation pipeline can effectively recover this structure. Our results demonstrate that automatic models can predict key dimensions, with prediction errors significantly correlating with human disagreement. By overcoming the constraints of single-score metrics, this work establishes a fine-grained, diagnostically valuable evaluation paradigm for open-ended generation tasks, offering deeper insights into model performance and human alignment.
This work addresses the high computational complexity of message-passing detectors in massive MIMO systems by proposing a low-complexity receiver framework based on orbital priors. By relaxing discrete symbol priors into mixed discrete–continuous densities and leveraging an orbital prior decomposition, the symbol distribution is compressed into 3L real-valued scalars, yielding three closed-form denoisers—OBD, OGD, and OPD—that reduce per-iteration complexity to O(1). Combining the Jacobi–Anger expansion, state evolution analysis, and optimal transport theory, the proposed method asymptotically achieves capacity (log₂M), eliminates error floors, attains an MMSE dimension of d = 1/2, and establishes a Wasserstein-distance bound linking constellation ring geometry to achievable rates.
研究通过优化SKILL文件来提高编码代理性能,使用合并的拉取请求作为更难任务,并基于代理表现评分,发现GEPA方法有效提升了文档质量。
该文介绍了一种使用大型语言模型辅助开发静态验证软件的工具Eiffel-tools,通过与语言服务器协议结合,并利用形式化验证器提高代码生成和修正的准确性。
本文针对异构敏捷地球观测卫星调度问题,提出了一种结合强化学习引导的进化策略优化框架,通过解码器和基于群体的搜索方法实现高效优化。
This study addresses the limitations of holistic scoring in metaphor explanation evaluation, which often overlooks quality structure and human disagreement. We propose a cognition-driven, six-dimensional assessment framework to capture these nuances. Through large-scale annotation and clustering analysis, we reveal the multidimensionality of explanation quality and systematic patterns of disagreement, validating that an automated evaluation pipeline can effectively recover this structure. Our results demonstrate that automatic models can predict key dimensions, with prediction errors significantly correlating with human disagreement. By overcoming the constraints of single-score metrics, this work establishes a fine-grained, diagnostically valuable evaluation paradigm for open-ended generation tasks, offering deeper insights into model performance and human alignment.
This work addresses the high computational complexity of message-passing detectors in massive MIMO systems by proposing a low-complexity receiver framework based on orbital priors. By relaxing discrete symbol priors into mixed discrete–continuous densities and leveraging an orbital prior decomposition, the symbol distribution is compressed into 3L real-valued scalars, yielding three closed-form denoisers—OBD, OGD, and OPD—that reduce per-iteration complexity to O(1). Combining the Jacobi–Anger expansion, state evolution analysis, and optimal transport theory, the proposed method asymptotically achieves capacity (log₂M), eliminates error floors, attains an MMSE dimension of d = 1/2, and establishes a Wasserstein-distance bound linking constellation ring geometry to achievable rates.