When to Review: Spaced Repetition for Continual Pre-Training of Language Models
研究解决了持续预训练中知识遗忘问题，通过引入基于认知科学的间隔重复训练方法来调度样本复习，提高旧知识保留和新知识获取。
0 citationsRead paper
Selected work
Representative Papers
EvoTS-Agent: A Self-Evolving LLM Agent for Financial Time Series Change Point Detection
为解决金融时间序列变点检测难题，提出EvoTS-Agent，通过验证指导的自我进化方法自动优化检测模型。
0 citationsRead paper
GNN-based Multi-Agent Control of Traffic Shockwaves in Sparse Vehicular Ad-hoc Networks
This study addresses the challenges posed by traffic shockwaves, which exacerbate congestion, reduce fuel efficiency, and elevate accident risks. Existing control strategies often rely on global traffic information, limiting their applicability in sparse vehicular ad hoc networks (VANETs). To overcome this limitation, this work proposes a decentralized cooperative control framework that integrates graph neural networks (GNNs) with multi-agent reinforcement learning (MARL), enabling connected automated vehicles to effectively suppress shockwave propagation using only local observations and interactions with neighboring vehicles. Notably, this approach is the first to incorporate GNNs into MARL under sparse VANET conditions, substantially enhancing practicality for early-stage deployment. Simulation results on a highway scenario with only 10% market penetration demonstrate up to an 80% reduction in shockwave propagation, confirming the method’s efficacy and scalability.
0 citationsRead paper
Dr. SHAP-AV: Decoding Relative Modality Contributions via Shapley Attribution in Audio-Visual Speech Recognition
This work addresses the unclear dynamics of how audio-visual speech recognition (AVSR) systems weigh contributions from audio and visual modalities under noisy conditions. The authors propose Dr. SHAP-AV, a novel framework that introduces Shapley values into AVSR for the first time, quantifying modality contributions across three dimensions: global importance, generation process, and temporal alignment. Their analysis reveals that AVSR models remain significantly reliant on audio even at low signal-to-noise ratios (SNR), that modality weights are primarily governed by SNR and evolve dynamically during decoding, and that temporal alignment mechanisms exhibit robustness in noisy environments. Extensive experiments across two benchmarks and six state-of-the-art models validate the effectiveness of Dr. SHAP-AV, offering a new tool for interpretability and diagnostic analysis in AVSR research.
0 citationsRead paper
Evaluating Performance Drift from Model Switching in Multi-Turn LLM Systems
This study addresses the performance degradation caused by context mismatch when switching large language models (LLMs) in multi-turn dialogues—a phenomenon termed "silent drift." The work presents the first systematic characterization of this issue and introduces a benchmark evaluation framework based on a switching matrix. By employing pairwise turn-level bootstrap confidence intervals, the authors quantify the impact of model switching on dialogue performance across datasets such as CoQA and Multi-IF. A key innovation lies in decomposing the drift into prefix influence and suffix sensitivity, revealing systematic robustness or vulnerability of models to non-self-generated context. Experiments demonstrate that a single switch can alter Multi-IF’s strict success rate by −8 to +13 percentage points and induce CoQA F1 fluctuations of up to ±4 points, offering both theoretical insights and practical tools for effective risk monitoring.
0 citationsRead paper
Recent publications
Latest Papers
When to Review: Spaced Repetition for Continual Pre-Training of Language Models
研究解决了持续预训练中知识遗忘问题，通过引入基于认知科学的间隔重复训练方法来调度样本复习，提高旧知识保留和新知识获取。
0 citationsRead paper
EvoTS-Agent: A Self-Evolving LLM Agent for Financial Time Series Change Point Detection
为解决金融时间序列变点检测难题，提出EvoTS-Agent，通过验证指导的自我进化方法自动优化检测模型。
0 citationsRead paper
GNN-based Multi-Agent Control of Traffic Shockwaves in Sparse Vehicular Ad-hoc Networks
This study addresses the challenges posed by traffic shockwaves, which exacerbate congestion, reduce fuel efficiency, and elevate accident risks. Existing control strategies often rely on global traffic information, limiting their applicability in sparse vehicular ad hoc networks (VANETs). To overcome this limitation, this work proposes a decentralized cooperative control framework that integrates graph neural networks (GNNs) with multi-agent reinforcement learning (MARL), enabling connected automated vehicles to effectively suppress shockwave propagation using only local observations and interactions with neighboring vehicles. Notably, this approach is the first to incorporate GNNs into MARL under sparse VANET conditions, substantially enhancing practicality for early-stage deployment. Simulation results on a highway scenario with only 10% market penetration demonstrate up to an 80% reduction in shockwave propagation, confirming the method’s efficacy and scalability.
0 citationsRead paper
Dr. SHAP-AV: Decoding Relative Modality Contributions via Shapley Attribution in Audio-Visual Speech Recognition
This work addresses the unclear dynamics of how audio-visual speech recognition (AVSR) systems weigh contributions from audio and visual modalities under noisy conditions. The authors propose Dr. SHAP-AV, a novel framework that introduces Shapley values into AVSR for the first time, quantifying modality contributions across three dimensions: global importance, generation process, and temporal alignment. Their analysis reveals that AVSR models remain significantly reliant on audio even at low signal-to-noise ratios (SNR), that modality weights are primarily governed by SNR and evolve dynamically during decoding, and that temporal alignment mechanisms exhibit robustness in noisy environments. Extensive experiments across two benchmarks and six state-of-the-art models validate the effectiveness of Dr. SHAP-AV, offering a new tool for interpretability and diagnostic analysis in AVSR research.
0 citationsRead paper
Evaluating Performance Drift from Model Switching in Multi-Turn LLM Systems
This study addresses the performance degradation caused by context mismatch when switching large language models (LLMs) in multi-turn dialogues—a phenomenon termed "silent drift." The work presents the first systematic characterization of this issue and introduces a benchmark evaluation framework based on a switching matrix. By employing pairwise turn-level bootstrap confidence intervals, the authors quantify the impact of model switching on dialogue performance across datasets such as CoQA and Multi-IF. A key innovation lies in decomposing the drift into prefix influence and suffix sensitivity, revealing systematic robustness or vulnerability of models to non-self-generated context. Experiments demonstrate that a single switch can alter Multi-IF’s strict success rate by −8 to +13 percentage points and induce CoQA F1 fluctuations of up to ±4 points, offering both theoretical insights and practical tools for effective risk monitoring.
0 citationsRead paper