Execution-Guided Code Generation
Overview
Execution-guided code generation uses evidence from running programs—such as test results, compiler errors, runtime traces, and rendered outputs—to help language models produce and repair code. Common approaches iteratively revise a candidate, generate or refine tests, select among candidates by comparing behavior, or train models with execution-based rewards and step-level feedback. These methods are evaluated on settings ranging from programming benchmarks such as HumanEval, MBPP, and LiveCodeBench to repository tasks, database queries, and generated interfaces. A key tension is that execution feedback can catch syntax and runtime failures but may be noisy, costly, or misleading when tests are incomplete or the model and verifier share a mistaken assumption.
Latest papers
67 papers8 papers in the last four weeks, up 100% on the four weeks before. 0.0% of all new papers.