Neuro-symbolic program synthesis for structured reasoning
aicpp is an experimental neuro-symbolic system that learns to search executable symbolic program spaces.
The project explores how neural representations can guide symbolic program synthesis while preserving executable, inspectable and verifiable solutions.
The idea
Given a task and a domain-specific language (DSL), aicpp learns to navigate a constrained program space rather than directly predicting the final output.
Problem
↓
Neural representation
↓
Program-space guidance
↓
Constrained program synthesis
↓
Executable symbolic program
↓
Verified solution
Current research
The project currently investigates:
- neural guidance of symbolic program search
- joint representations of visual tasks and programs
- constrained autoregressive program generation
- executable DSLs
- program embeddings and latent structure
- experience reuse across sequential synthesis tasks
ARC-AGI-2
aicpp is currently evaluated on ARC-AGI-2.
View the ARC-AGI-2 experiments
Research results
Architecture
Publications
Publications and research material