KANACKI Symbolic Regression
Perfect scores on Feynman (100/100) · Nguyen (12/12) · Generalization (26/26) · No evolutionary search, no neural components
What is KANACKI?
We present KANACKI, a symbolic regression framework that achieves state-of-the-art performance on established benchmarks without relying on:
- evolutionary search
- grammar-based expression generation
- neural proposal models
- computer algebra engines
- variable-separation procedures
- domain-specific decomposition strategies
To our knowledge, this is the first symbolic regression system that achieves competitive benchmark performance without any of these components.
How It Works
KANACKI operates using only deterministic expression expansion, numerical evaluation, and utility-guided beam search augmented by a linear regression composition layer. The complete implementation is released as a single Python file with no external dependencies beyond NumPy.
Benchmark Performance
On the Feynman benchmark of 100 physics equations, KANACKI achieves a perfect solution rate of 100/100, matching AI Feynman while requiring none of its physics-inspired heuristics or neural components.
On the Nguyen benchmark (12/12) and Nguyen+Noise benchmark (12/12), KANACKI achieves perfect recovery.
On a held-out generalization benchmark of 26 novel equations spanning polynomial, trigonometric, exponential, logarithmic, and composite families, KANACKI again achieves 100% exact recovery.
Robustness & Ablation
Ablation studies demonstrate that the framework's performance is robust to 4× beam width reduction and that linear regression composition is the sole non-trivial contributor beyond beam search itself.
The complete implementation is released as a single Python file with no external dependencies beyond NumPy — making it easy to integrate, audit, and extend.