Zhihao Lin

Zhihao Lin, Yongtao Wang, Jinhe Zhang et al.

Dynamic neural network is an emerging research topic in deep learning. With adaptive inference, dynamic models can achieve remarkable accuracy and computational efficiency. However, it is challenging to design a powerful dynamic detector, because of no suitable dynamic architecture and exiting criterion for object detection. To tackle these difficulties, we propose a dynamic framework for object detection, named DynamicDet. Firstly, we carefully design a dynamic architecture based on the nature of the object detection task. Then, we propose an adaptive router to analyze the multi-scale information and to decide the inference route automatically. We also present a novel optimization strategy with an exiting criterion based on the detection losses for our dynamic detectors. Last, we present a variable-speed inference strategy, which helps to realize a wide range of accuracy-speed trade-offs with only one dynamic detector. Extensive experiments conducted on the COCO benchmark demonstrate that the proposed DynamicDet achieves new state-of-the-art accuracy-speed trade-offs. For instance, with comparable accuracy, the inference speed of our dynamic detector Dy-YOLOv7-W6 surpasses YOLOv7-E6 by 12%, YOLOv7-D6 by 17%, and YOLOv7-E6E by 39%. The code is available at https://github.com/VDIGPKU/DynamicDet.

Zhihao Lin, Ziqi Zhu, Hao Huang et al.

Literary translation poses unique challenges due to the scarcity of high-quality annotated data and the need to balance expression fluency with literary effect. We present a multi-aspect iterative refinement framework that generates high-quality translation references and preference data through specialized LLM translators, each targeting a distinct quality dimension. We leverage the generated data for supervised fine-tuning and reinforcement learning. Experiments show that our generated references outperform the original ground truth for SFT by 8.65 CEA100 points. For reinforcement learning, we find that DPO leads to performance degradation in this setting, while leveraging an explicit reward model for GRPO yields an additional 1.51 point improvement. We attribute this to the stability of two-stage training and GRPO's online exploration capability. Our resulting models, LitMT-8B and LitMT-14B, achieve 67.25 and 69.07 CEA100 respectively on the MetaphorTrans English-to-Chinese literary translation benchmark, competitive with Claude Sonnet 4.5 at 68.43, and demonstrate strong generalization to out-of-domain literary work (i.e., O. Henry).

Zhihao Lin, Wei Ma, Mingyi Zhou et al.

In recent years, Large Language Models (LLMs) have gained widespread use, raising concerns about their security. Traditional jailbreak attacks, which often rely on the model internal information or have limitations when exploring the unsafe behavior of the victim model, limiting their reducing their general applicability. In this paper, we introduce PathSeeker, a novel black-box jailbreak method, which is inspired by the game of rats escaping a maze. We think that each LLM has its unique "security maze", and attackers attempt to find the exit learning from the received feedback and their accumulated experience to compromise the target LLM's security defences. Our approach leverages multi-agent reinforcement learning, where smaller models collaborate to guide the main LLM in performing mutation operations to achieve the attack objectives. By progressively modifying inputs based on the model's feedback, our system induces richer, harmful responses. During our manual attempts to perform jailbreak attacks, we found that the vocabulary of the response of the target model gradually became richer and eventually produced harmful responses. Based on the observation, we also introduce a reward mechanism that exploits the expansion of vocabulary richness in LLM responses to weaken security constraints. Our method outperforms five state-of-the-art attack techniques when tested across 13 commercial and open-source LLMs, achieving high attack success rates, especially in strongly aligned commercial models like GPT-4o-mini, Claude-3.5, and GLM-4-air with strong safety alignment. This study aims to improve the understanding of LLM security vulnerabilities and we hope that this sturdy can contribute to the development of more robust defenses.

Zhihao Lin, Yongtao Wang, Jinhe Zhang et al.

Binary Neural Networks (BNNs) have gained extensive attention for their superior inferencing efficiency and compression ratio compared to traditional full-precision networks. However, due to the unique characteristics of BNNs, designing a powerful binary architecture is challenging and often requires significant manpower. A promising solution is to utilize Neural Architecture Search (NAS) to assist in designing BNNs, but current NAS methods for BNNs are relatively straightforward and leave a performance gap between the searched models and manually designed ones. To address this gap, we propose a novel neural architecture search scheme for binary neural networks, named NAS-BNN. We first carefully design a search space based on the unique characteristics of BNNs. Then, we present three training strategies, which significantly enhance the training of supernet and boost the performance of all subnets. Our discovered binary model family outperforms previous BNNs for a wide range of operations (OPs) from 20M to 200M. For instance, we achieve 68.20% top-1 accuracy on ImageNet with only 57M OPs. In addition, we validate the transferability of these searched BNNs on the object detection task, and our binary detectors with the searched BNNs achieve a novel state-of-the-art result, e.g., 31.6% mAP with 370M OPs, on MS COCO dataset. The source code and models will be released at https://github.com/VDIGPKU/NAS-BNN.

Qi Zhang, Zhihao Lin, Arnoud Visser

The ICRA conference is celebrating its $40^{th}$ anniversary in Rotterdam in September 2024, with as highlight the Happy Birthday ICRA Party at the iconic Holland America Line Cruise Terminal. One month later the IROS conference will take place, which will include the Earth Rover Challenge. In this challenge open-world autonomous navigation models are studied truly open-world settings. As part of the Earth Rover Challenge several real-world navigation sets in several cities world-wide, like Auckland, Australia and Wuhan, China. The only dataset recorded in the Netherlands is the small village Oudewater. The proposal is to record a dataset with the robot used in the Earth Rover Challenge in Rotterdam, in front of the Holland America Line Cruise Terminal, before the festivities of the Happy Birthday ICRA Party start. See: https://github.com/SlamMate/vSLAM-on-FrodoBots-2K

Zhihao Lin, Zhen Tian, Jianglin Lan et al.

Safety and efficiency are crucial for autonomous driving in roundabouts, especially mixed traffic with both autonomous vehicles (AVs) and human-driven vehicles. This paper presents a learning-based algorithm that promotes safe and efficient driving across varying roundabout traffic conditions. A deep Q-learning network is used to learn optimal strategies in complex multi-vehicle roundabout scenarios, while a Kolmogorov-Arnold Network (KAN) improves the AVs' environmental understanding. To further enhance safety, an action inspector filters unsafe actions, and a route planner optimizes driving efficiency. Moreover, model predictive control ensures stability and precision in execution. Experimental results demonstrate that the proposed system consistently outperforms state-of-the-art methods, achieving fewer collisions, reduced travel time, and stable training with smooth reward convergence.

Zhihao Lin, Wei Ma, Tao Lin et al.

Large Language Models (LLMs) have become instrumental in advancing software engineering (SE) tasks, showcasing their efficacy in code understanding and beyond. Like traditional SE tools, open-source collaboration is key in realising the excellent products. However, with AI models, the essential need is in data. The collaboration of these AI-based SE models hinges on maximising the sources of high-quality data. However, data especially of high quality, often holds commercial or sensitive value, making it less accessible for open-source AI-based SE projects. This reality presents a significant barrier to the development and enhancement of AI-based SE tools within the software engineering community. Therefore, researchers need to find solutions for enabling open-source AI-based SE models to tap into resources by different organisations. Addressing this challenge, our position paper investigates one solution to facilitate access to diverse organizational resources for open-source AI models, ensuring privacy and commercial sensitivities are respected. We introduce a governance framework centered on federated learning (FL), designed to foster the joint development and maintenance of open-source AI code models while safeguarding data privacy and security. Additionally, we present guidelines for developers on AI-based SE tool collaboration, covering data requirements, model architecture, updating strategies, and version control. Given the significant influence of data characteristics on FL, our research examines the effect of code data heterogeneity on FL performance.

Zhihao Lin, Lin Wu, Zhen Tian et al.

Exploration in reinforcement learning remains a critical challenge, as naive entropy maximization often results in high variance and inefficient policy updates. We introduce \textbf{PrefPoE}, a novel \textit{Preference-Product-of-Experts} framework that performs intelligent, advantage-guided exploration via the first principled application of product-of-experts (PoE) fusion for single-task exploration-exploitation balancing. By training a preference network to concentrate probability mass on high-advantage actions and fusing it with the main policy through PoE, PrefPoE creates a \textbf{soft trust region} that stabilizes policy updates while maintaining targeted exploration. Across diverse control tasks spanning both continuous and discrete action spaces, PrefPoE demonstrates consistent improvements: +321\% on HalfCheetah-v4 (1276~$\rightarrow$~5375), +69\% on Ant-v4, +276\% on LunarLander-v2, with consistently enhanced training stability and sample efficiency. Unlike standard PPO, which suffers from entropy collapse, PrefPoE sustains adaptive exploration through its unique dynamics, thereby preventing premature convergence and enabling superior performance. Our results establish that learning \textit{where to explore} through advantage-guided preferences is as crucial as learning how to act, offering a general framework for enhancing policy gradient methods across the full spectrum of reinforcement learning domains. Code and pretrained models are available in supplementary materials.

Zhihao Lin

Standard stochastic policies for continuous control often rely on ad-hoc boundary-enforcing transformations (e.g., tanh) which can distort the underlying optimization landscape and introduce gradient pathologies. While alternative parameterizations on the unit manifold (e.g., directional distributions) are theoretically appealing, their computational complexity (often requiring special functions or rejection sampling) has limited their practical use. We propose a novel, computationally efficient action generation paradigm that preserves the structural benefits of operating on a unit manifold. Our method decomposes the action into a deterministic directional vector and a learnable concentration scalar, enabling efficient interpolation between the target direction and uniform noise on the unit manifold. This design can reduce policy head parameters by nearly 50\% (from $2d$ to $d+1$) and maintains a simple $O(d)$ sampling complexity, avoiding costly sampling procedures. Empirically, our method matches or exceeds state-of-the-art methods on standard continuous control benchmarks, with significant improvements (e.g., +37.6\% and +112\%) on high-dimensional locomotion tasks. Ablation studies confirm that both the unit-norm normalization and the adaptive concentration mechanism are essential to the method's success. These findings suggest that robust, efficient control can be achieved by explicitly respecting the structure of bounded action spaces, rather than relying on complex, unbounded distributions. Code is available in supplementary materials.

Zhensu Sun, Zhihao Lin, Zhi Chen et al.

Current LLM-based coding agents follow a serial execution paradigm: the model first generates the complete code, then invokes an interpreter to execute it. This sequential workflow leaves the executor idle during generation and the generator idle during execution, resulting in unnecessary end-to-end latency. We observe that, unlike human developers, LLMs produce code tokens sequentially without revision, making it possible to execute code as it is being generated. We formalize this parallel execution paradigm, modeling it as a three-stage pipeline of generation, detection, and execution, and derive closed-form latency bounds that characterize its speedup potential and operating regimes. We then present Eager, a concrete implementation featuring AST-based chunking, dynamic batching with gated execution, and early error interruption. We evaluate Eager across four benchmarks, seven LLMs, and three execution environments. Results show that Eager reduces the non-overlapped execution latency by up to 99.9% and the end-to-end latency by up to 55% across seven LLMs and four benchmarks.

Zhen Tian, Zhihao Lin, Dezong Zhao et al.

Autonomous vehicles (AVs) can significantly promote the advances in road transport mobility in terms of safety, reliability, and decarbonization. However, ensuring safety and efficiency in interactive during within dynamic and diverse environments is still a primary barrier to large-scale AV adoption. In recent years, deep reinforcement learning (DRL) has emerged as an advanced AI-based approach, enabling AVs to learn decision-making strategies adaptively from data and interactions. DRL strategies are better suited than traditional rule-based methods for handling complex, dynamic, and unpredictable driving environments due to their adaptivity. However, varying driving scenarios present distinct challenges, such as avoiding obstacles on highways and reaching specific exits at intersections, requiring different scenario-specific decision-making algorithms. Many DRL algorithms have been proposed in interactive decision-making. However, a rationale review of these DRL algorithms across various scenarios is lacking. Therefore, a comprehensive evaluation is essential to assess these algorithms from multiple perspectives, including those of vehicle users and vehicle manufacturers. This survey reviews the application of DRL algorithms in autonomous driving across typical scenarios, summarizing road features and recent advancements. The scenarios include highways, on-ramp merging, roundabouts, and unsignalized intersections. Furthermore, DRL-based algorithms are evaluated based on five rationale criteria: driving safety, driving efficiency, training efficiency, unselfishness, and interpretability (DDTUI). Each criterion of DDTUI is specifically analyzed in relation to the reviewed algorithms. Finally, the challenges for future DRL-based decision-making algorithms are summarized.

Wei Ma, Shangqing Liu, Zhihao Lin et al.

Large language models~(LLMs) demonstrate significant potential to revolutionize software engineering (SE) by exhibiting outstanding performance in SE tasks such as code and document generation. However, the high reliability and risk control requirements in software engineering raise concerns about the lack of interpretability of LLMs. To address this concern, we conducted a study to evaluate the capabilities of LLMs and their limitations for code analysis in SE. We break down the abilities needed for artificial intelligence~(AI) models to address SE tasks related to code analysis into three categories: 1) syntax understanding, 2) static behavior understanding, and 3) dynamic behavior understanding. Our investigation focused on the ability of LLMs to comprehend code syntax and semantic structures, which include abstract syntax trees (AST), control flow graphs (CFG), and call graphs (CG). We employed four state-of-the-art foundational models, GPT4, GPT3.5, StarCoder and CodeLlama-13b-instruct. We assessed the performance of LLMs on cross-language tasks involving C, Java, Python, and Solidity. Our findings revealed that while LLMs have a talent for understanding code syntax, they struggle with comprehending code semantics, particularly dynamic semantics. We conclude that LLMs possess capabilities similar to an Abstract Syntax Tree (AST) parser, demonstrating initial competencies in static code analysis. Furthermore, our study highlights that LLMs are susceptible to hallucinations when interpreting code semantic structures and fabricating nonexistent facts. These results indicate the need to explore methods to verify the correctness of LLM output to ensure its dependability in SE. More importantly, our study provides an initial answer to why the codes generated by LLM are usually syntax-correct but vulnerable.

Shengzhe Wang, Ling Wang, Zhihao Lin et al.

Existing Deep Neural Nets on crops growth prediction mostly rely on availability of a large amount of data. In practice, it is difficult to collect enough high-quality data to utilize the full potential of these deep learning models. In this paper, we construct an innovative network architecture based on domain adaptation learning to predict crops growth curves with limited available crop data. This network architecture overcomes the challenge of data availability by incorporating generated data from the developed crops simulation model. We are the first to use the temporal convolution filters as the backbone to construct a domain adaptation network architecture which is suitable for deep learning regression models with very limited training data of the target domain. We conduct experiments to test the performance of the network and compare our proposed architecture with other state-of-the-art methods, including a recent LSTM-based domain adaptation network architecture. The results show that the proposed temporal convolution-based network architecture outperforms all benchmarks not only in accuracy but also in model size and convergence rate.