Computer Science M.S.

• Built a supervised in-context learning framework that trains sequence models on optimization trajectories to predict topology-aware step sizes and doubly stochastic mixing matrices. On sparse graphs with edge probability p=0.3, the method reduced iterations by up to 200 compared with FDGM, DIGing, and related baselines.
• Extended trajectory-supervised hyperparameter prediction to time-varying graphs, enabling online adaptation of communication weights during topology evolution and supporting optimization runs beyond 500 iterations.
• Replaced heuristic edge selection and neighbor weighting with trajectory-conditioned prediction, reducing convergence iterations by up to 300 under Gossip and Broadcast protocols on Census and Digits datasets.

• Expanded the framework to broader optimization algorithm discovery tasks and evaluated its out-of-distribution generalization across four datasets against dozens of baselines, outperforming competing methods on most problems.
• Applied the framework to distributed optimization algorithm discovery on static and time-varying undirected and directed graphs, successfully reproducing classical algorithms and validating module necessity through ablation studies.
• Explored transfer to reinforcement learning evaluation-function discovery and achieved effective results on synthetic datasets, demonstrating the framework's transfer potential.

• Deployed and integrated Affine XiaoZhi and OpenClaw to support scalable, modular AI pipeline systems.
• Debugged cross-layer issues spanning software, runtime, and hardware, improving system stability and deployment reliability.

github.com/PetrichorHlacyon

计算机科学 硕士

• 基于监督式上下文学习框架，利用优化轨迹训练序列模型，预测拓扑最优步长与双随机混合矩阵。实验表明，在稀疏图（p=0.3）场景下，与FDGM和DIGing等基线方法相比，迭代次数最多可减少200次。
• 将轨迹监督的超参数预测方法扩展至时变图场景，实现通信权重在拓扑演化过程中的在线自适应调整，支持复杂优化任务迭代至500次以上。
• 通过轨迹条件化预测替代传统启发式边选择与邻居加权策略，在Census和Digits数据集的Gossip与Broadcast协议中，收敛迭代次数最多减少300次。

• 拓展该框架至更广泛的优化算法发现任务，在包含数十个基线的对照实验中，使用4个数据集验证了框架在分布外任务中的泛化能力，相较于其他基线在多数问题上表现更优。
• 应用于静态/时变无向图和有向图的分布式优化算法发现，成功复现经典算法，并通过消融实验证实各模块对算法发现的必要性。
• 探索框架在强化学习评估函数发现中的应用，在虚拟数据集上取得有效结果，展示了技术迁移潜力。

• 部署并集成 Affine XiaoZhi 与 OpenClaw 框架，以支持可扩展、模块化的 AI 流水线系统。
• 调试软件、运行时与硬件层面的跨层系统问题，提高系统稳定性与部署可靠性。

github.com/PetrichorHlacyon