Self-Modifying Program Synthesis via Online Library Evolution: A Mechanistic Approach to Few-Shot Abstract Reasoning with Aerospace Validation

I developed and evaluated SMPMA (Self-Modifying Program Synthesis with Adaptive Library Evolution), a system for few-shot abstract reasoning. The system’s key feature is that its capability space expands during problem-solving through online library evolution. SMPMA implements a three-stage loop. First, it synthesizes programs using beam search with current primitives and learned abstractions. Second, it mines abstractions from successful programs, including both fixed sequences and parameterized templates. Third, it estimates cross-task utility to promote valuable abstractions into the DSL. I evaluate on two benchmarks. On ARC-AGI (400 training tasks, 120 evaluation tasks), SMPMA achieves 2.5% strict accuracy with 47 learned abstractions. On AeroSynth (a controlled synthetic benchmark), it achieves 62% accuracy. These results validate the mechanism on compositional tasks. I also present a spacecraft rendezvous/docking autonomy prototype. This implements the same research principle of abstraction over an interpretable, verifiable structure. The spacecraft system achieves 100% success across 20 random seeds using LQR control with docking mode and safety shield (compared to 70% without shield). It also achieves 79% fuel reduction using sampling-based model predictive control (\(3.59\pm 0.12\) vs \(17.13\pm 0.35\) m/s \(\Delta v\))