SInQ: Synthetic Data Generation for Code Reasoning in LLMs

A groundbreaking technique has been developed by researchers to produce synthetic training data aimed at enhancing Large Language Models' (LLMs) capabilities in intricate code reasoning tasks. This approach, termed the Semantic Inequivalence Game (SInQ), utilizes two agents: a generator that produces program variants differing semantically from actual programming tasks, and an evaluator that detects instances where these variants exhibit different behaviors. These agents engage in a semi-adversarial training process, improving their grasp of code logic. The method was tested across various benchmarks for code generation and understanding, including cross-language vulnerability detection, yielding improved performance in identifying vulnerabilities. Detailed findings are available in a paper on arXiv (2505.03818), which includes a replace-cross announcement. This method tackles the challenge of sourcing training examples necessary for LLMs to comprehend program semantics, essential for complex tasks beyond routine coding.