Scalable code generation and analysis RL environment scalable for large language models (LLMs) training

Studying existing environments for procedural reasoning and code generation, such as Reasoning Gym (Shen et al., Reasoning Gym: Procedurally Generated Reasoning Environments for Pretraining and Evaluating LLMs, 2024), DELTA-Code (Sun et al., DELTA-Code: A Benchmark for RL-based Algorithm Learning in Code LLMs, ICLR 2025), StepCoder (Dou et al., StepCoder: Improving LLM-Based Code Generation via Step-wise RL, 2024), and RLEF (Gehring et al., Reinforcement Learning from Execution Feedback for Code Generation, NeurIPS 2024).

Identifying gaps in existing approaches, in particular the lack of procedurally generated static analysis tasks (e.g., call-graph construction, error detection, performance profiling).

Proposing new generators for Python code tasks that produce infinite, verifiable problem instances, with evaluation grounded in unit tests, execution traces, or static analyzers.

Implement procedural generators for code tasks (Python focus), inspired by TinyPy-style grammar synthesis (Naïr et al., Curriculum Learning for Code Execution with Synthetic Python Programs, 2023) and code perturbation methods

Integrate these generators in a Gym-style RL environment with automatic scoring and curriculum scheduling (easy → hard).

Explore a novel task family (e.g., static code reasoning, code completion, debugging, refactoring, performance analysis, or multi-file repository puzzles) and evaluate whether LLMs can improve through curriculum-based RL.