✦ ɪᴍʙᴜᴇ ✦

*When the world stopped, I started compiling.*

> *他会拿起一块蓝色玻璃，透过它看花园，花园里的沙地和路径会变成一种灰烬般的颜色，天空则变得异常深邃，仿佛热带的天空。*
> — 《说吧，记忆》，弗拉基米尔·纳博科夫

🔬 Recent Research & Publications

My research focuses on bridging the gap between human intuition and machine scale, with a strong emphasis on Deep Learning Theory, LLM Reasoning, Long-Context Processing, and Quantitative Finance.

📂 Research by Domain

📐 Deep Learning Theory & Scaling Laws

• Data Scaling as Progressive Coverage of a Predictive Contribution Spectrum

* *Core Contribution:* Real-data scaling laws are driven by progressively covering a latent predictive state spectrum rather than just token frequencies. By representing corpora as suffix-automata and using a global-KL spectrum, effective truncation ranks reliably predict excess loss across training sizes ($R^2 \approx 0.96$).

• A Formal Kinetic Theory for Zeroth-Order Newton Dynamics

* *Core Contribution:* Develops a kinetic framework for Z-O Newton methods, providing a Stein-corrected Hessian estimator and exposing the curvature-variance trade-off.

• Reconstructing High-Resolution Hyperparameter Loss Landscapes

* *Core Contribution:* Frames hyperparameter tuning as a landscape reconstruction problem, using active surrogate modeling to find robust, generalizable minima.

🧠 LLM Reasoning & Alignment

• Inverting the Search Dynamics: LLMs as Semantic Leaders in MCTS

* *Core Contribution:* Proposes *Leader-Follower MCTS*, where the LLM steers search with macro-actions, improving performance on GSM8K, MATH, and HumanEval.

• The Statistical Illusion of Rejection Sampling in LLMs

* *Core Contribution:* Bridges the gap between heuristic truncation in LLM sampling and true mathematical alignment, revealing statistical biases.

• Expected Value Alignment for Generative Reward Modeling

* *Core Contribution:* Introduces *EVA*, a reward modeling paradigm for theorem proving that extracts continuous scores from discrete token distributions.

• Soft-NBCE: Entropy-Weighted Chunk Fusion for Long-Context Decoding

* *Core Contribution:* Presents *Soft-NBCE*, which replaces hard chunk selection with soft fusion, improving reasoning while maintaining memory efficiency.

📈 Quantitative Finance & Applied Systems

• Deep Learning under Continuous Distribution Shift for Quant Finance

* *Core Contribution:* Formulates a non-stationary NTK and spectral tracking SDE to model DL performance under persistent market distribution shifts.

• AdaPrecise: A Task-Agnostic Dynamic Precision Routing Framework

* *Core Contribution:* A Gumbel-Softmax based framework for dynamic precision routing that optimizes model efficiency for inference on edge devices.

🚗 Vision-Language-Action (VLA) & Embodied AI

• Lagrange: An Open-Vocabulary, Energy-Based Sparse Framework for Driving

* *Core Contribution:* A sparse, energy-based framework for autonomous driving that uses VLMs for class-agnostic perception and Lagrangian action minimization.

🌍 Beyond the Papers: My Multiverse of Engineering & Finance

• 📈 Quantitative Finance: I actively manage ~5 Million CNY in quantitative funds, integrating modern CS and deep learning into strategies to generate alpha.
• ⚙️ Systems & DevOps: An advocate for Clean Code & TDD. I've achieved C10K via kernel tuning/IO multiplexing and ran CPU-based IDC services with a 1:20 overselling ratio. I'm proficient in K8s (Helm, Prometheus, Grafana, ELK), and have improved resource utilization by 17% through HPA and Limit/Request tuning.
• 🌐 Frontend & UX: With a deep focus on aesthetics and HCI, I leverage React, Vue, and Ionic to transform complex backend logic into elegant user experiences. An excellent system needs both robustness in algorithms and poetry in its UI.
• 🛡️ CTF & CP: I'm active in XCTF (Crypto & Web) with contributions in problem-setting and write-ups, alongside a brief but intense stint in Competitive Programming.

🤖 Open Source & Trained Models

• Socrates-nano: Open-sourced the complete LLM codebase including pre-training, data synthesis pipelines, post-training, and test-time scaling.
• Socrates-embedding: A next-gen embedding model that outperforms an 83× larger parameter counterpart, achieving SOTA accuracy under identical budgets.
• RWKV-7-Prover-1.5B: A formal math model leveraging RWKV-7 & Condor-inspired data synthesis for high-precision Lean 4 auto-formalization.
• LPR-Oracle: A forecasting model for China’s Loan Prime Rate (LPR) in financial markets.
• Thales: Thales is an interpretable, physics-informed deep learning surrogate model for ultra-fast, arbitrage-free option pricing and AI-decoded risk reporting.

📊 GitHub Analytics

⚙️ Runtime Configuration

💻 System.Current() -> struct AboutMe

C
#include 

struct Skills {
    struct Languages {
        const char* proficient[6];
        const char* familiar[4];
        const char* exploring[5];
    } languages;

    struct Frontend {
        const char* frameworks_libraries[4];
        const char* styling[5];
        const char* state_management[3];
        const char* tools[3];
    } frontend;

    struct Backend {
        const char* frameworks_runtime[4];
        const char* databases[4];
        const char* orms[3];
        const char* apis[2];
    } backend;
    
    struct DataScience {
        const char* libraries[2];
        const char* tools[2];
    } data_science;

    struct DevOpsAndCloud {
        const char* containerization[2];
        const char* ci_cd[1];
        const char* cloud_platforms[3];
    } devops;

    struct ToolsAndEnvironment {
        const char* version_control[2];
        const char* editors_ides[3];
        const char* operating_systems[3];
        const char* design_tools[2];
    } tools;
};

struct AboutMe {
    const char* name;
    const int age;
    const char* gender;
    const char* interests[5];
    struct Skills skills;
};

struct AboutMe me = {
    .name = "imbue",
    .age = 15,
    .gender = "Female",
    .interests = {
        "LLM & Theoretical Machine Learning",
        "Quantitative Finance",
        "Full-Stack & Cloud Native",
        "Competitive Programming",
        "Cryptography & Infosec"
    },
    .skills = {
        .languages = {
            .proficient = { "C++", "Python", "JavaScript", "TypeScript", "HTML5", "CSS3" },
            .familiar = { "Rust", "Go", "Java", "SQL" },
            .exploring = { "Haskell", "Lisp", "C", "x86 Assembly", "QASM" }
        },
        .frontend = {
            .frameworks_libraries = { "React", "Next.js", "Vue.js", "Svelte" },
            .styling = { "Tailwind CSS", "Sass/SCSS", "Bootstrap", "Material-UI", "Styled-components" },
            .state_management = { "Redux", "Zustand", "Pinia" },
            .tools = { "Vite", "Webpack", "Babel" }
        },
        .backend = {
            .frameworks_runtime = { "Node.js", "Express.js", "FastAPI (Python)", "Actix Web (Rust)" },
            .databases = { "PostgreSQL", "MySQL", "MongoDB", "Redis" },
            .orms = { "Prisma", "SQLAlchemy (Python)", "Sequelize" },
            .apis = { "RESTful APIs", "GraphQL (Apollo)" }
        },
        .data_science = {
            .libraries = { "PyTorch", "Python (NumPy, Pandas, Scikit-learn)", "R (ggplot2)" },
            .tools = { "Jupyter Notebook", "SQL" }
        },
        .devops = {
            .containerization = { "Docker", "Kubernetes (Helm)" },
            .ci_cd = { "GitHub Actions" },
            .cloud_platforms = { "Vercel", "AWS", "Prometheus/Grafana", "ELK" }
        },
        .tools = {
            .version_control = { "Git", "GitHub" },
            .editors_ides = { "VS Code", "Neovim", "JetBrains IDEs" },
            .operating_systems = { "Linux (Ubuntu, CentOS, Manjaro, Rocky Linux)", "Windows (WSL2)", "macOS" },
            .design_tools = { "Figma", "Adobe XD" }
        }
    }
};

---

*Life is a stochastic process; optimize for the long tail.*

---

荣誉勋章


Collapse