ReaComp: Compiling LLM Reasoning into Symbolic Solvers for Efficient Program Synthesis

ReaComp introduces a method to compile reasoning traces from large language models into reusable symbolic program synthesizers that eliminate runtime LLM calls. The approach achieves 91.3% accuracy on benchmark tasks while reducing token usage by 78%, demonstrating that neuro-symbolic hybrid systems can outperform pure LLM inference on complex program synthesis problems.

ReaComp addresses a fundamental limitation in LLM-based program synthesis: inefficiency and unreliability on computationally complex tasks. By extracting reasoning patterns from LLM traces and compiling them into symbolic solvers, the research demonstrates that expensive inference can be replaced with deterministic computation at test time. This shift from continuous inference to discrete symbolic execution represents an important efficiency frontier in AI systems.

The neuro-symbolic hybrid approach reflects broader trends in machine learning toward complementary architectures. Rather than treating LLMs as universal solvers, this work recognizes that different problem classes benefit from different computational substrates. Symbolic solvers excel at exhaustive search over constrained domains, while LLMs provide valuable pattern recognition during the training phase. The 78% reduction in token usage while improving accuracy highlights the economic advantage of this decomposition.

For developers and AI practitioners, ReaComp suggests practical pathways to deploy LLM capabilities more cost-effectively in production systems. The zero-shot transfer to linguistics tasks indicates the approach generalizes beyond synthetic benchmarks, enabling domain adaptation without retraining. This creates opportunities for organizations to reduce inference costs substantially while maintaining or improving accuracy on specialized tasks.

The implications extend to AI infrastructure planning and model economics. As token costs drive operational expenses, systems that compile reasoning into reusable symbolic components could become increasingly valuable. Future work likely explores automated trace extraction and solver induction pipelines that minimize manual intervention in the compilation process.

• →Symbolic solvers compiled from LLM reasoning traces achieve 91.3% accuracy on program synthesis benchmarks without runtime LLM calls.
• →Neuro-symbolic hybrid approach reduces token usage by 78% while improving accuracy from 68.4% to 85.8% on hard benchmark tasks.
• →Induced solvers transfer zero-shot to real-world tasks like linguistics, reaching 80.1% accuracy and recovering interpretable linguistic rules.
• →Symbolic solvers provide better Pareto efficiency than per-instance LLM agents by amortizing construction costs across many executions.
• →The method demonstrates that reasoning traces can be compiled into reusable, domain-general solvers for complex combinatorial search problems.