TerraFormer: Automated Infrastructure-as-Code with LLMs Fine-Tuned via Policy-Guided Verifier Feedback

Automating Infrastructure-as-Code (IaC) is challenging, and large language models (LLMs) often produce incorrect configurations from natural language (NL). We present TerraFormer, a neuro-symbolic framework for IaC generation and mutation that combines supervised fine-tuning with verifier-guided reinforcement learning, using formal verification tools to provide feedback on syntax, deployability, and policy compliance. We curate two large, high-quality NL-to-IaC datasets, TF-Gen (152k instances) and TF-Mutn (52k instances), via multi-stage verification and iterative LLM self-correction. Evaluations against 17 state-of-the-art LLMs, including ~50x larger models like Sonnet 3.7, DeepSeek-R1, and GPT-4.1, show that TerraFormer improves correctness over its base LLM by 15.94% on IaC-Eval, 11.65% on TF-Gen (Test), and 19.60% on TF-Mutn (Test). It outperforms larger models on both TF-Gen (Test) and TF-Mutn (Test), ranks third on IaC-Eval, and achieves top best-practices and security compliance.

REDO: Execution-Free Runtime Error Detection for COding Agents

As LLM-based agents exhibit exceptional capabilities in addressing complex problems, there is a growing focus on developing coding agents to tackle increasingly sophisticated tasks. Despite their promising performance, these coding agents often produce programs or modifications that contain runtime errors, which can cause code failures and are difficult for static analysis tools to detect. Enhancing the ability of coding agents to statically identify such errors could significantly improve their overall performance. In this work, we introduce Execution-free Runtime Error Detection for COding Agents (REDO), a method that integrates LLMs with static analysis tools to detect runtime errors for coding agents, without code execution. Additionally, we propose a benchmark task, SWE-Bench-Error-Detection (SWEDE), based on SWE-Bench (lite), to evaluate error detection in repository-level problems with complex external dependencies. Finally, through both quantitative and qualitative analyses across various error detection tasks, we demonstrate that REDO outperforms current state-of-the-art methods by achieving a 11.0% higher accuracy and 9.1% higher weighted F1 score; and provide insights into the advantages of incorporating LLMs for error detection.