Inference-Time Code Selection via Symbolic Equivalence Partitioning

Researchers propose Symbolic Equivalence Partitioning, a novel inference-time selection method for code generation that uses symbolic execution and SMT constraints to identify correct solutions without expensive external verifiers. The approach improves accuracy on HumanEval+ by 10.3% and on LiveCodeBench by 17.1% at N=10 without requiring additional LLM inference.

This research addresses a critical bottleneck in scaling code generation with Large Language Models: selecting correct solutions from multiple candidates efficiently. Traditional "Best-of-N" approaches rely on external verifiers—such as unit tests or execution traces—that are computationally expensive and sometimes unreliable. The proposed Symbolic Equivalence Partitioning method leverages symbolic execution to group candidate programs by their semantic behavior, then selects a representative from the largest functional partition. This approach elegantly sidesteps the verifier problem by assuming that the most common correct behavior across multiple independently generated solutions is likely correct.

The integration of Satisfiability Modulo Theories constraints during symbolic execution represents a pragmatic engineering choice. By encoding domain-specific constraints, the method reduces path explosion—a fundamental problem in symbolic execution—and prevents the system from exploring invalid input spaces. This constrains the symbolic search space to semantically meaningful regions, improving both efficiency and accuracy.

For the AI infrastructure market, this method has immediate implications. Code generation tools used by developers and enterprises can dramatically improve reliability without deploying expensive verification infrastructure or incurring additional computational costs beyond the initial N candidate generations. The results—10% improvements on HumanEval+ and 17% on LiveCodeBench—suggest meaningful real-world productivity gains. This research may influence how companies design inference pipelines, potentially reducing operational costs while improving output quality. The technique is particularly valuable for resource-constrained deployments where external verifiers are prohibitively expensive.

• →Symbolic Equivalence Partitioning selects correct code solutions by grouping candidates by semantic behavior rather than relying on expensive external verifiers.
• →SMT constraints during symbolic execution reduce path explosion and improve efficiency without requiring additional LLM inference.
• →Accuracy improvements of 10.3% on HumanEval+ and 17.1% on LiveCodeBench demonstrate practical value for code generation scaling.
• →The method reduces computational overhead by leveraging N already-generated candidates rather than requiring new model calls.
• →Integration of domain-specific constraints during symbolic execution improves both selection accuracy and algorithmic performance.