VRAXION INSTNCT: Exploration of a Gradient-Free Self-Connected Graph Neural Network Architecture

This article introduces the INSTNCT architecture developed by the VRAXION team, a neural network that challenges the traditional backpropagation paradigm. Its core feature is that it does not rely on gradient descent to optimize fixed topology; instead, it learns by dynamically changing its own directed graph structure, using innovative mechanisms such as phased self-connection and scout-first search. This article will discuss aspects including background, mechanisms, validation, implementation, and contributions.

In the field of deep learning, backpropagation has long been the core training method. However, the VRAXION team challenges this paradigm by developing the INSTNCT architecture—which completely abandons gradient computation and achieves learning by changing the network's topological structure, forming a stark contrast to the traditional forward propagation + backpropagation parameter adjustment approach.

INSTNCT's core hypothesis is that "reasoning emerges as a fixed point of destructive interference": signals enter the cyclic matrix, incompatible paths cancel each other out, and the surviving pattern is the reasoning result. Its architecture uses a phased self-connection mechanism:

1. Scout-first search: First perform a low-cost full-signal detection to rank potential parent nodes, then conduct an exhaustive ternary search to balance efficiency and cost;
2. Bias-free threshold neurons: Directly store the form where the dot product ≥ threshold, reducing the search space and improving interpretability.

VRAXION uses three labels to mark the maturity of work:

• Current mainline: Delivered code on the main branch;
• Validated finding: Supported by experiments but not in the canonical path;
• Experimental branch: Actively under development (not validated by default). The code-first principle ensures that the main branch code is the source of truth. The current public version v5.0.0-beta.9 implements D10u phase state-anchored connection search and passes the adversarial 16k gate test (H=384) with 30 random seeds, but it is not a mainline alternative—it is only a research checkpoint.

The INSTNCT tech stack has evolved from Python to Rust:

• Early Python research code is archived at archives/python-research-20260420;
• The current mainline is the Rust implementation instnct-core/examples/neuron_grower.rs;
• The Python deployment SDK (in the Python/ directory) is still maintained, with a pure NumPy implementation that has no ML framework dependencies;
• Next milestone: Byte/Opcode v1 (1-byte +4 opcodes →1-byte converter), which is the main threshold for entering the public beta.

VRAXION values reproducibility:

• The canonical grower contract is defined in docs/GROWER_RUN_CONTRACT.md, requiring canonical runs to produce reproducible evidence packages and recoverable states;
• Validation commands are provided:
  • cargo test -p instnct-core: Validate Rust library compilation and tests;
  • python tools/run_grower_regression.py: Validate reproducibility of grower regression packages;
  • python tools/run_byte_opcode_acceptance.py: Validate Byte/Opcode v1 export/overload;
  • python tools/check_public_surface.py: Validate consistency between documentation and code.

Contributions:

1. Freedom from gradient dependency: Avoids traditional issues like gradient vanishing/explosion;
2. Dynamic topology: Network structure is a learning result rather than a hyperparameter;
3. Interpretability potential: The interference-based reasoning mechanism is more interpretable than black-box weights. Limitations:

• In research preview phase, not for commercial use;
• Only validated under H=384 configuration;
• Has not reached mainline version maturity;
• Learning efficiency and scale scalability need further validation.

VRAXION's roadmap points to the "grower-based v5.0.0 public beta". If the Byte/Opcode v1 translator is frozen and reaches production quality, INSTNCT may become an important alternative direction. For the research community, it provides a thought experiment of "gradient-independent structural self-organization learning", inspiring new algorithmic ideas. Even if it does not become mainstream, it adds new possibilities to the field of deep learning and is worth continuing to follow.