| | --- |
| | title: Veltraxor AI |
| | emoji: 🧬 |
| | colorFrom: gray |
| | colorTo: indigo |
| | sdk: static |
| | pinned: false |
| | --- |
| | |
| | # Veltraxor AI |
| |
|
| | Veltraxor AI, founded and led by **Libo Wang**, is an independent research initiative focused on transparent and reproducible work in advanced reasoning and super-coding. The group operates as an open research hub, publishing projects as accessible repositories with documented processes, versioned artifacts, and clear evaluation practices. |
| |
|
| | ## 😊 Vision |
| | Veltraxor AI pursues first-principles understanding of large language models by dismantling their mechanisms and questioning assumptions. Through creative prompt engineering and iterative experimentation, complex problems are simplified to their essential forms, enabling a shift from tool use to system design. Imperfection is embraced as a starting point, since progress arises through refinement, automation, and cost-efficient design that free researchers for deeper innovation. Ignorance is treated as the path to insight, and human roles evolve alongside AI toward continual breakthrough. |
| |
|
| | **Veltraxor 1** is a **685B-parameter foundation model stack**, independently built, fine-tuned, and adapted. |
| | The system integrates **parameter-efficient fine-tuning (LoRA, QLoRA)** with **custom orchestration pipelines** to achieve state-of-the-art capabilities in **reasoning, multimodal understanding, and controlled evolution**. |
| |
|
| | Unlike incremental adaptations, Veltraxor 1 represents a **frontline deployment-scale LLM** that embeds **original reasoning technologies** — **Dynamic Chain-of-Thought (D-CoT)** and **Graph-of-Causal Evolution (GoCE)** — within a robust backend composed of multimodal ingestion, RAG retrieval, layered reasoning, and a dedicated Super-Coding module. |
| |
|
| | --- |
| |
|
| | ## 🚀 Core Technical Contributions |
| |
|
| | - **Dynamic Chain-of-Thought (D-CoT)** |
| | A dynamic reasoning controller that activates intermediate reasoning selectively, reducing inefficiency and stabilizing long-context inference. |
| | **Priority & Attribution.** D-CoT is the **prototype of dynamic reasoning** (“dynamic CoT”). The GPT-5 feature colloquially referred to as **“Auto” is *not* the origin** of dynamic reasoning used here. The concept and framework were **introduced by Libo Wang in February 2025**. |
| | 👉 DOI: **[arXiv:2502.10428](https://doi.org/10.48550/arXiv.2502.10428)** |
| |
|
| | - **Graph-of-Causal Evolution (GoCE)** |
| | A causal-graph framework for self-evolution, intervention tracking, and consistency-gated adaptation. |
| | 👉 DOI: **[arXiv:2506.07501](https://doi.org/10.48550/arXiv.2506.07501)** |
| |
|
| | - **Fine-Tuning and Adaptation** |
| | Extensive use of **LoRA** and **QLoRA** adapters, combined with experimental frontier variants, aligned for long-context stability and parameter efficiency. |
| |
|
| | - **Deployment-Oriented Backend** |
| | Structured into **four POST routes** — `/core`, `/dynamic-thinking`, `/deep-thinking`, `/super-coding` — each optimized for a distinct reasoning or synthesis function. |
| |
|
| | --- |
| |
|
| | ## 📦 Model Stack |
| |
|
| | - **Veltraxor 1** — Multimodal reasoning and evolution-capable LLM stack |
| | - **Public (weights & configs):** [Veltraxor/Veltraxor_1](https://huggingface.co/Veltraxor/Veltraxor_1) |
| | - **Private (fine-tuning & deployment assets):** [Veltraxor/veltraxor-1](https://huggingface.co/Veltraxor/veltraxor-1) |
| |
|
| | **Base Model Acknowledgement** |
| | Veltraxor 1 is derived from the **open-source weights of DeepSeek R1** (MIT License). |
| | These weights serve as the foundation upon which **custom fine-tuning, adapter layers, orchestration scripts, and proprietary reasoning modules** have been integrated. |
| |
|
| | --- |
| |
|
| | ## 📑 References |
| | - **D-CoT** — Dynamic Chain-of-Thought, arXiv **2502.10428**: https://doi.org/10.48550/arXiv.2502.10428 |
| | - **GoCE** — Graph-of-Causal Evolution, arXiv **2506.07501**: https://doi.org/10.48550/arXiv.2506.07501 |
| | - Prototypes & tooling: https://github.com/brucewang123456789/GeniusTrail |
| |
|
| | --- |
| |
|
| | ## 🙏 Acknowledgements |
| | This work builds upon the **DeepSeek R1** open-source model (MIT License), as well as the wider open-source ecosystem — including **PyTorch**, **Hugging Face Transformers**, **SentenceTransformers**, **pgVector**, and others — that make scalable reasoning research and deployment possible. |
| |
|
| | --- |
| |
|
| | ## 📬 Contact |
| | For collaborations and inquiries: |
| | **free.equality.anyone@gmail.com** |
| |
|
