Soft Symbol Grounding for Prototypical Concepts
该研究提出Soft-PNet模型，通过原型分布和KL散度目标解决神经符号模型中的推理捷径问题，无需手动设计任务特定损失函数。
0 citationsRead paper
Selected work
Representative Papers
Generalized Hamming Weights of AJ-Gorenstein One-Point Codes
研究AJ-Gorenstein曲线一点码的广义汉明重量，通过构建零图并利用扭曲和Wei对偶性，确定了超过一半的广义权重位置。
0 citationsRead paper
Measuring Similarity between Artistic and AI Generated Images using Siamese Neural Networks
研究使用冻结CLIP编码器的Siamese神经网络和余弦相似度来量化原作与AI生成图像间的相似性，解决AI艺术潜在抄袭问题。
0 citationsRead paper
Leveraging Remote Traffic Data for Local Air Pollutant Estimation: A Scenario-Based Machine Learning Study Across London Monitoring Sites
研究利用远程交通数据和机器学习模型（如随机森林、XGBoost等）估计伦敦各站点的空气污染物浓度，比较了不同预测情景下的模型表现。
0 citationsRead paper
Why Summaries Turn Neutral: Policy Attribution for Sentiment Drift in Reinforcement Learning from Human Feedback
This study addresses sentiment drift and excessive neutralization in RLHF-based summarization by proposing a strategy attribution framework that identifies the model’s low-risk preference as the underlying cause. Building on this insight, we introduce a sentiment-aware regularization technique that integrates gradient and logit decomposition to effectively mitigate drift, while also validating the cross-lingual universality of these drift patterns. Experimental results demonstrate that our approach reduces sentiment drift by 18%–22% and significantly improves sentiment fidelity without compromising summary quality. These findings establish a novel paradigm for fine-grained sentiment control during the alignment stage of large language model training.
0 citationsRead paper
Recent publications
Latest Papers
Soft Symbol Grounding for Prototypical Concepts
该研究提出Soft-PNet模型，通过原型分布和KL散度目标解决神经符号模型中的推理捷径问题，无需手动设计任务特定损失函数。
0 citationsRead paper
Generalized Hamming Weights of AJ-Gorenstein One-Point Codes
研究AJ-Gorenstein曲线一点码的广义汉明重量，通过构建零图并利用扭曲和Wei对偶性，确定了超过一半的广义权重位置。
0 citationsRead paper
Measuring Similarity between Artistic and AI Generated Images using Siamese Neural Networks
研究使用冻结CLIP编码器的Siamese神经网络和余弦相似度来量化原作与AI生成图像间的相似性，解决AI艺术潜在抄袭问题。
0 citationsRead paper
Leveraging Remote Traffic Data for Local Air Pollutant Estimation: A Scenario-Based Machine Learning Study Across London Monitoring Sites
研究利用远程交通数据和机器学习模型（如随机森林、XGBoost等）估计伦敦各站点的空气污染物浓度，比较了不同预测情景下的模型表现。
0 citationsRead paper
Why Summaries Turn Neutral: Policy Attribution for Sentiment Drift in Reinforcement Learning from Human Feedback
This study addresses sentiment drift and excessive neutralization in RLHF-based summarization by proposing a strategy attribution framework that identifies the model’s low-risk preference as the underlying cause. Building on this insight, we introduce a sentiment-aware regularization technique that integrates gradient and logit decomposition to effectively mitigate drift, while also validating the cross-lingual universality of these drift patterns. Experimental results demonstrate that our approach reduces sentiment drift by 18%–22% and significantly improves sentiment fidelity without compromising summary quality. These findings establish a novel paradigm for fine-grained sentiment control during the alignment stage of large language model training.
0 citationsRead paper