Dual-System Architecture: A Large Language Model Enhancement Scheme Without Modifying Base Model Weights

The Dual-System Architecture introduced in this article is an innovative LLM enhancement scheme. Its core lies in adding a "geometric sidecar" module to achieve functions like uncensored generation, structured mathematical reasoning, multi-user isolation, and continuous learning without modifying the base model weights. This architecture treats the frozen base model as "System 1" (fast intuition) and the sidecar module as "System 2" (slow reasoning). It supports independent training iterations, avoids the risk of model degradation caused by traditional fine-tuning, and is compatible with various mainstream LLM architectures (e.g., Qwen2.5-3B, Llama-3.1-8B, etc.).

Traditional LLM enhancement usually relies on fine-tuning or continued pre-training, but it has problems such as high computational cost, difficulty in rollback, and potential damage to original capabilities. The Dual-System Architecture proposes a "geometric sidecar" design, which enhances frozen base models by adding trainable modules, solving the above pain points and providing a new path for LLM capability expansion.

The core design of the Dual-System Architecture is the "System 1 + System 2" model: System 1 is the frozen base LLM, and System 2 is the geometric sidecar module. The sidecar module includes several key components:

1. Diffusion Planner: Based on DDIM and adaptive layer normalization, it converts input tokens into high-dimensional latent planning representations;
2. Geometric Processor: A 4-layer Transformer architecture that performs geometric transformations on latent representations;
3. Fiber Bundle: A per-user personalization mechanism based on principal fiber bundle theory, ensuring user isolation (cross-user cos_sim ≥ 0.9999);
4. EBM Judge: Based on an energy-based model, it identifies factual hallucinations and style mismatches;
5. Cognitive Router: Routes gradients via the Kappa gating mechanism to mitigate catastrophic forgetting in continuous learning.

Validation of Key Capabilities:

• Uncensored Generation: Using the FailSpy differential mean method to extract and project the "rejection direction", the rejection rate dropped from about 80% to 0%, while the difference between 6 benchmark tests (ARC-E, ARC-C, etc.) and the baseline was ≤0.3 percentage points;
• Multi-User Isolation: The cos_sim of output changes from per-user style correction was 0.997, and cross-user isolation cos_sim ≥0.9999;
• Continuous Learning: The EBM Judge performs token-level credit assignment, the Cognitive Router automatically routes gradients, and the BCH integrator merges cross-session perturbations;
• TurboKV Cache Compression: In 4-bit mode, memory usage was reduced by 3.9x (from 896MB to 232MB for 8K context).

Hardware and Deployment Support:

• Hardware Performance: When running Qwen2.5-3B-Instruct on RTX4060Ti, the peak memory usage is 3.4GB, and the generation speed is 36 tokens per second; dual-GPU sharding deployment is supported;
• Web Dashboard: Provides functions such as neural terminal, tensor telemetry, generation control, feedback loop, and checkpoint management;
• API Services: OpenAI-compatible API server that supports streaming/non-streaming generation and feedback endpoints, enabling multi-concurrent inference and continuous learning updates with exclusive GPU access.

Open Source Ecosystem:

• Open-sourced under the Apache 2.0 license, including training pipelines, benchmark tools, and unit tests;
• Pre-trained ablation models and sidecar checkpoints have been uploaded to HuggingFace;
• Includes the M-A-K-E-R multi-role audit framework for autonomous security analysis of decentralized protocols.

The Dual-System Architecture represents a new paradigm for LLM enhancement: it achieves capability expansion through mathematically rigorous附加 modules without modifying the base model. Its advantages include reducing experimental iteration costs, supporting multi-tenant deployment, continuous learning, and personalized services. For researchers and developers focusing on local AI deployment, model security, and efficient inference, this project has significant exploration value.