Distilling Neuro-Symbolic Programs into 3D Multi-modal LLMs

Researchers introduce APEIRIA, a neuro-symbolic 3D multi-modal language model that combines the interpretability of symbolic AI with the flexibility of modern LLMs for 3D spatial reasoning. The system uses a three-stage curriculum to distill reasoning patterns from symbolic programs into natural language chain-of-thought, achieving performance competitive with state-of-the-art models while maintaining transparent, modular reasoning.

APEIRIA addresses a longstanding limitation in AI systems handling 3D spatial understanding. Traditional neuro-symbolic approaches offer interpretable, verifiable reasoning through compositional programs but struggle with real-world complexity and open-vocabulary concepts. Conversely, end-to-end 3D multi-modal LLMs handle natural language fluently and scale to diverse concepts but operate as black boxes, making their spatial reasoning opaque and difficult to verify—a critical drawback for applications requiring trustworthiness.

The research represents a meaningful convergence in AI methodology. Rather than choosing between interpretability and capability, APEIRIA transfers reasoning patterns from symbolic systems into LLM architectures through a structured curriculum. The three-stage approach—perception alignment, chain-of-thought supervised fine-tuning from symbolic traces, and reinforcement learning on nested instructions—systematically builds reasoning capabilities while preserving modularity. This design allows swapping perception or planning components independently, maintaining advantages of neuro-symbolic systems.

For the AI industry, this work signals growing recognition that interpretability and scale need not be mutually exclusive. Industries requiring explainable AI—robotics, autonomous systems, scientific discovery—stand to benefit from models balancing reasoning transparency with language flexibility. The approach validates that symbolic knowledge can guide neural scaling rather than being entirely abandoned. However, the practical deployment impact remains unclear without broader evaluation across domain-specific tasks and real-world deployment scenarios.

• →APEIRIA bridges neuro-symbolic and end-to-end LLM paradigms by distilling reasoning patterns into explainable language-based chains-of-thought.
• →The three-stage curriculum progressively builds 3D spatial reasoning from perception alignment through reinforcement learning on complex instructions.
• →System maintains modular interchangeability of components while matching state-of-the-art performance, preserving transparency advantages of symbolic methods.
• →Approach validates that symbolic program traces can effectively guide neural language model training for interpretable spatial reasoning.
• →Represents broader trend toward hybrid AI architectures combining interpretability and scalability rather than treating them as opposing forces.