Suan: Rectifying Direct Preference Safety Alignment in Large Language Models
本文提出Suan算法，通过在梯度级别直接优化偏好，解决大型语言模型中安全对齐问题，同时保持响应的实用性。
0 citationsRead paper
Selected work
Representative Papers
Test-time adaptation for speech enhancement with an autoregressive speech prior
研究提出一种单语句测试时自适应方法，利用自回归语音先验调整预训练的语音增强模型，以改善声学条件不匹配下的语音质量。
0 citationsRead paper
Reasoning for Social Audio-Visual Question Answering: Where Do We Stand?
This work addresses critical limitations in existing audio-visual social understanding benchmarks, which suffer from high noise levels and poorly designed questions, while complex reasoning approaches often incur substantial costs with marginal gains. The authors systematically evaluate the reasoning capabilities of multimodal large language models on social audio-visual question answering, introducing IntentBench-Prime—a high-quality benchmark constructed through rigorous data cleaning—and comparing diverse training strategies. Their findings reveal that a simple vanilla supervised fine-tuning (SFT) baseline matches or surpasses state-of-the-art complex methods across three benchmarks. Notably, using only textual captions achieves performance comparable to full video inputs, suggesting that linguistic modalities encode strong social priors. The study further proposes a cost-effective evaluation paradigm and publicly releases the denoised IntentBench-Prime benchmark to support future research.
0 citationsRead paper
When Imbalance Comes Twice: Active Learning under Simulated Class Imbalance and Label Shift in Binary Semantic Segmentation
This study addresses the degradation of active learning performance in binary semantic segmentation caused by the coexistence of class imbalance and label shift. For the first time, it systematically simulates both challenges jointly on open-source datasets to evaluate the effectiveness of three active learning strategies: random sampling, entropy maximization, and core-set selection. Experimental results demonstrate that entropy-based and core-set methods remain robust under severe class imbalance; however, strong label shift significantly impairs their performance. By revealing distinct behavioral patterns of these strategies under compound distribution shifts, this work provides critical insights for deploying active learning in real-world scenarios where multiple data biases may co-occur.
0 citationsRead paper
Recent publications
Latest Papers
Suan: Rectifying Direct Preference Safety Alignment in Large Language Models
本文提出Suan算法，通过在梯度级别直接优化偏好，解决大型语言模型中安全对齐问题，同时保持响应的实用性。
0 citationsRead paper
Test-time adaptation for speech enhancement with an autoregressive speech prior
研究提出一种单语句测试时自适应方法，利用自回归语音先验调整预训练的语音增强模型，以改善声学条件不匹配下的语音质量。
0 citationsRead paper
Reasoning for Social Audio-Visual Question Answering: Where Do We Stand?
This work addresses critical limitations in existing audio-visual social understanding benchmarks, which suffer from high noise levels and poorly designed questions, while complex reasoning approaches often incur substantial costs with marginal gains. The authors systematically evaluate the reasoning capabilities of multimodal large language models on social audio-visual question answering, introducing IntentBench-Prime—a high-quality benchmark constructed through rigorous data cleaning—and comparing diverse training strategies. Their findings reveal that a simple vanilla supervised fine-tuning (SFT) baseline matches or surpasses state-of-the-art complex methods across three benchmarks. Notably, using only textual captions achieves performance comparable to full video inputs, suggesting that linguistic modalities encode strong social priors. The study further proposes a cost-effective evaluation paradigm and publicly releases the denoised IntentBench-Prime benchmark to support future research.
0 citationsRead paper
When Imbalance Comes Twice: Active Learning under Simulated Class Imbalance and Label Shift in Binary Semantic Segmentation
This study addresses the degradation of active learning performance in binary semantic segmentation caused by the coexistence of class imbalance and label shift. For the first time, it systematically simulates both challenges jointly on open-source datasets to evaluate the effectiveness of three active learning strategies: random sampling, entropy maximization, and core-set selection. Experimental results demonstrate that entropy-based and core-set methods remain robust under severe class imbalance; however, strong label shift significantly impairs their performance. By revealing distinct behavioral patterns of these strategies under compound distribution shifts, this work provides critical insights for deploying active learning in real-world scenarios where multiple data biases may co-occur.
0 citationsRead paper