reinforcement learning

VietLegalLM: Progressive Legal Expertise Through Synthetic Comprehension and Reinforcement Learning

Legal AI systems require accuracy and verifiable reasoning, yet under-resourced languages lack the specialized models needed to meet these standards. This challenge is particularly acute for statute-based civil law systems like Vietnam’s, where the core task is interpreting and applying codified statutes rather than matching legal precedents. We address this gap by introducing a comprehensive framework for developing reliable legal AI under resource constraints. First, we present vilaw-bench, a novel evaluation benchmark tailored for Vietnamese statute-based legal reasoning that assesses cognitive abilities from basic knowledge retrieval to complex legal interpretation. Second, we propose a multi-stage training framework that progressively builds legal expertise through four phases: focused foundational learning on core legal texts, intensive comprehension practice using large-scale synthetic question-answer data, targeted skill acquisition through supervised fine-tuning, and response quality refinement using Group Relative Policy Optimization (GRPO). We implement this framework to develop VietLegalLM, training on Qwen3-1.7B-Base and Qwen3-4B-Base models. Our systematic evaluation reveals that synthetic comprehension practice produces the largest single-phase improvements in legal reasoning capabilities, while GRPO efficiently refines reasoning structure with minimal training steps. The complete sequential training approach achieves significant performance gains on vilaw-bench compared with baseline models. Our ablation studies demonstrate that practitioners can make informed tradeoffs between comprehensive training and computational efficiency: direct GRPO after pre-training offers a viable alternative to full instruction tuning when resources are limited. We release vilaw-bench, VietLegalLM, and our training framework as open-source resources, providing a reproducible roadmap for developing legal AI in other under-resourced, statute-based legal systems.

基于多智能体强化学习的软件开发流程优化

This paper investigates the application of Multi-Agent Reinforcement Learning (MARL) to optimize software development processes. Traditional software development methodologies often struggle with adaptability and efficiency, particularly in complex projects. This research proposes a novel approach leveraging MARL to dynamically adjust and refine the development workflow. The core idea involves modeling the software development process as a multi-agent system, where each agent represents a distinct stage or activity. These agents learn optimal strategies through interaction and reward signals, leading to improved development efficiency and quality. We present a framework for formulating this problem, detailing the agent architecture, state space, action space, and reward function. The effectiveness of the MARL approach is demonstrated through a theoretical analysis and conceptual design, highlighting its potential to overcome limitations of static, rule-based methodologies. Future work will focus on implementing and testing this framework within a simulated software development environment.

An adaptive unscented Kalman Filter with nonlinear model predictive control based on reinforcement learning for the induction motor control in electric vehicles

Mitigating perception-induced false safety in USV navigation: Memory- and perception-credibility-aware adaptive-risk distributional reinforcement learning

Cooperative optimization of energy and thermal management in PHEV using a time-correlated Laplacian noise-based reinforcement learning framework

Multi-task deep reinforcement learning for service-oriented post-flood recovery scheduling of large urban metro systems

Geometric Topology Optimization Algorithms

This paper introduces Geometric Topology Optimization Algorithms (GTOA) – a novel approach that integrates deep learning with topology optimization to accelerate complex structural simulations and design. We propose a novel "拓扑感知" algorithm utilizing Generative Adversarial Networks (GANs) to learn the geometric structure of target shapes and combine it with reinforcement learning to automatically generate optimization paths. The core mechanism lies in automating the exploration process, overcoming the limitations of traditional methods. The paper details the algorithm's architecture, training process, and validation results, demonstrating its potential for efficient and intelligent geometric structure design. The new algorithm promises to significantly reduce design time and improve the accuracy of optimized structures.

Adaptive Learning-Based Robot Motion Planning

This paper presents a novel approach to robot motion planning based on adaptive learning. Traditional robot motion planning methods often struggle to cope with dynamic and uncertain environments. This research leverages reinforcement learning (RL) to develop a system capable of autonomously learning optimal motion strategies and adapting to real-time changes in the environment and task requirements. The core of the system is a dynamically adjusted trajectory generation process, allowing the robot to navigate complex scenarios with greater flexibility and efficiency. The system is trained using an RL algorithm, and the learned policy is then utilized for real-time trajectory optimization. This approach demonstrates the potential for significantly enhancing the adaptability and intelligence of robots in various applications. The system's ability to learn and adapt provides a robust solution for dynamic environments, addressing limitations inherent in traditional planning techniques. ---

Explainable Reinforcement Learning with Causal Inference

This paper explores a novel approach to reinforcement learning (RL) that integrates causal inference to enhance both the reliability and interpretability of decision-making processes. Traditional RL algorithms often operate as "black boxes," lacking transparency into the reasoning behind their actions. This research addresses this limitation by leveraging causal relationships within an environment to guide the learning process. We propose a framework where causal graphs are utilized to understand the underlying dynamics of the environment, informing the selection of optimal policies in an RL agent. Furthermore, we employ explainable AI (XAI) techniques to provide insights into the agent's decision-making rationale. The core claim is that by combining causal reasoning with RL and XAI, we can significantly improve the robustness and trustworthiness of the learned policies, while simultaneously increasing their interpretability. The proposed method aims to provide a more reliable and understandable system than standard RL approaches, particularly in complex and uncertain environments.

Dynamic Semantic Network Generator

This paper introduces a Dynamic Semantic Network Generator (DSNG), a novel system designed to automatically construct and maintain semantic networks representing complex, dynamic systems. The core of the DSNG is a reinforcement learning (RL) framework that adapts the network's structure and connections in response to changes in system states and external events. Unlike traditional static semantic networks, the DSNG allows for real-time semantic understanding and evolution, offering a more accurate and flexible representation of complex systems. The system utilizes a state representation, action space, and reward function to learn optimal network configurations. The key innovation lies in the dynamic adjustment mechanism, leveraging RL to continuously refine the network's topology and weights, mirroring the evolving relationships within the system under observation. This approach addresses the limitations of static models, providing a robust solution for analyzing and understanding systems where relationships are not fixed but constantly shifting. The system's performance is evaluated through simulations, demonstrating its ability to capture and adapt to complex temporal dependencies.

Dynamic Context-Aware Neural Architecture Search (DCA-NAS)

This paper introduces Dynamic Context-Aware Neural Architecture Search (DCA-NAS), a novel approach to neural architecture search that addresses the limitations of traditional NAS methods by incorporating dynamic contextual information during the search process. The core claim of DCA-NAS is that real-time analysis of intermediate representations generated during training—such as activation values and gradient information—can dynamically adjust the exploration strategy of the NAS algorithm, thereby accelerating the discovery of optimal architectures. DCA-NAS combines generative NAS techniques (e.g., evolutionary algorithms, reinforcement learning) with search-based NAS (e.g., differentiable architecture search) through an integrated "context module." This module learns to identify the contextual factors most influential on architecture selection, considering not only the statistical properties of the searched architecture but also the statistical properties of intermediate representations, training loss gradients, and learning rates. Based on this contextual understanding, the module dynamically adjusts the search strategy of the generative component or the search space of the search-based component. Furthermore, DCA-NAS divides the search process into multiple "sub-search" stages, each optimized for a specific context. Experimental results demonstrate DCA-NAS's ability to achieve superior architecture quality and search efficiency compared to static NAS methods, particularly in scenarios with diverse tasks and datasets.

Neural Network Dynamics: Topology Self-Organization

This paper investigates the potential of dynamically self-organizing neural networks inspired by biological systems. Traditional static neural networks often struggle with adaptability and efficiency in complex environments. We propose a novel approach that leverages reinforcement learning to dynamically adjust connection strengths and incorporates probabilistic mechanisms for synaptic generation and removal, mirroring the plasticity observed in biological neural networks. The network topology itself evolves over time, driven by the reinforcement learning process, leading to the formation of more effective connection patterns. Our simulations demonstrate that this dynamic topology self-organization significantly enhances learning and inference capabilities compared to static networks. The core claim is that by simulating dynamic connections and synaptic changes, we can achieve more efficient and adaptive learning and reasoning. The central mechanism utilizes reinforcement learning to adjust connection strengths based on network output success rates and introduces probabilistic synaptic generation and deletion to mimic biological synaptic plasticity. The network topology evolves over time, creating optimized connection patterns.