Rui Zou

Yin Luo, Qingchao Kong, Nan Xu et al.

As the latest advancements in natural language processing, large language models (LLMs) have achieved human-level language understanding and generation abilities in many real-world tasks, and even have been regarded as a potential path to the artificial general intelligence. To better facilitate research on LLMs, many open-source LLMs, such as Llama 2 and Falcon, have recently been proposed and gained comparable performances to proprietary models. However, these models are primarily designed for English scenarios and exhibit poor performances in Chinese contexts. In this technical report, we propose YAYI 2, including both base and chat models, with 30 billion parameters. YAYI 2 is pre-trained from scratch on a multilingual corpus which contains 2.65 trillion tokens filtered by our pre-training data processing pipeline. The base model is aligned with human values through supervised fine-tuning with millions of instructions and reinforcement learning from human feedback. Extensive experiments on multiple benchmarks, such as MMLU and CMMLU, consistently demonstrate that the proposed YAYI 2 outperforms other similar sized open-source models.

Changjun He, Li Zhang, Qi Liu et al.

This paper investigates the transient synchronization stability in power systems hybridized with virtual synchronous generators (VSGs) and synchronous generators (SGs). A relative swing equation model is established to capture the transient synchronization dynamics between the VSG and the SG. Based on this model, both static and dynamic characteristics are systematically analyzed, and a quantitative stability level index is derived to elucidate the underlying stability mechanism. Then, two fundamental inertia matching principles are identified. First, a new instability mechanism induced by improper inertia matching between the VSG and the SG is revealed. It is identified that increasing the VSG's inertia does not monotonically improve transient stability, as commonly presumed. Instead, an optimal inertia matching constant exists that maximizes stability performance. Second, the influence of the VSG share on the synchronization stability is discovered to be strongly influenced by the matching between the VSG's inertia level and its voltage strength (i.e., output impedance). To achieve reliable and robust synchronization stability, proper coordination between the VSG's inertia and virtual impedance is essential. Finally, a coordinated stabilization strategy based on inertia matching and virtual impedance adjustment is proposed to enhance transient synchronization stability performance while suppressing fault current. Simulations conducted on a two-machine system and the IEEE 39-bus system validate the theoretical findings and demonstrate the effectiveness of the proposed strategy.

Rui Zou, Mengqi Wei, Jintian Feng et al.

In recent years, large language models have shown exceptional performance in fulfilling diverse human needs. However, their training data can introduce harmful content, underscoring the necessity for robust value alignment. Mainstream methods, which depend on feedback learning and supervised training, are resource-intensive and may constrain the full potential of the models. Multi-Agent Debate (MAD) offers a more efficient and innovative solution by enabling the generation of reliable answers through agent interactions. To apply MAD to value alignment, we examine the relationship between the helpfulness and harmlessness of debate outcomes and individual responses, and propose a MAD based framework Gradual Vigilance and Interval Communication (GVIC). GVIC allows agents to assess risks with varying levels of vigilance and to exchange diverse information through interval communication. We theoretically prove that GVIC optimizes debate efficiency while reducing communication overhead. Experimental results demonstrate that GVIC consistently outperforms baseline methods across various tasks and datasets, particularly excelling in harmfulness mitigation and fraud prevention. Additionally, GVIC exhibits strong adaptability across different base model sizes, including both unaligned and aligned models, and across various task types.

Rui Zou

Ensemble learning boosts performance by aggregating predictions from multiple base learners. A core challenge is balancing individual learner accuracy with diversity. Traditional methods like Bagging and Boosting promote diversity through randomness but lack precise control over the accuracy-diversity trade-off. Negative Correlation Learning (NCL) introduces a penalty to manage this trade-off but suffers from loose theoretical bounds and limited adjustment range. To overcome these limitations, we propose a novel framework called Self-Error Adjustment (SEA), which decomposes ensemble errors into two distinct components: individual performance terms, representing the self-error of each base learner, and diversity terms, reflecting interactions among learners. This decomposition allows us to introduce an adjustable parameter into the loss function, offering precise control over the contribution of each component, thus enabling finer regulation of ensemble performance. Compared to NCL and its variants, SEA provides a broader range of effective adjustments and more consistent changes in diversity. Furthermore, we establish tighter theoretical bounds for adjustable ensemble methods and validate them through empirical experiments. Experimental results on several public regression and classification datasets demonstrate that SEA consistently outperforms baseline methods across all tasks. Ablation studies confirm that SEA offers more flexible adjustment capabilities and superior performance in fine-tuning strategies.

Rui Zou, Mengqi Wei, Yutao Zhu et al.

Large Language Models (LLMs) excel in reasoning and generation across domains, but still struggle with identifying and diagnosing complex errors. This stems mainly from training objectives that prioritize correct answers, limiting exposure to and learning from errors. While recent studies have begun to address this by introducing error signals, most rely on shallow, static errors, restricting improvement in deep diagnostic ability. To overcome this, we propose Hide and Seek Game (HSG), a dynamic adversarial framework for error generation and diagnosis, and evaluate it on mathematical problem-solving. HSG involves two adversarial roles: Sneaky, which "hides" by generating subtle, deceptive reasoning errors, and Diagnosis, which "seeks" to accurately detect them. Through adversarial co-evolution, both error stealth and diagnostic precision are enhanced. Experiments on several math reasoning tasks show that HSG significantly boosts error diagnosis, achieving 16.8\%--31.4\% higher accuracy than baselines like GPT-4o. We also release a challenging dataset of deceptive errors and diagnostic annotations as a benchmark for future research.

Guillermo J. Laguna, Rui Zou, Sourabh Bhattacharya

We investigate a variation of the art gallery problem in which a team of mobile guards tries to track an unpredictable intruder in a simply-connected polygonal environment. In this work, we use the deployment strategy for diagonal guards originally proposed in [1]. The guards are confined to move along the diagonals of a polygon and the intruder can move freely within the environment. We define critical regions to generate event-triggered strategies for the guards. We design a hybrid automaton based on the critical regions to model the tracking problem. Based on reachability analysis, we provide necessary and sufficient conditions for tracking in terms of the maximal controlled invariant set of the hybrid system. We express these conditions in terms of the critical curves to find sufficient conditions for n/4 guards to track the mobile intruder using the reachability analysis.