Selective Left-Shift: Turning Test-Time Compute and Difficulty-based Curation into Training Data for Low-Resource Code Generation

How can we improve code generation performance for low-resource programming languages (LRPLs) without relying on expensive inference-time scaling or massive, human-annotated training datasets? The authors address the trilemma of data scarcity, high computational costs, and the failure of reinforcement learning from scratch in small language models (SLMs).

The authors propose a three-phase pipeline that decouples syntax acquisition from algorithmic reasoning:

1. Offline Data Synthesis (Left-Shift): Instead of performing iterative refinement during deployment, the authors use an offline engine that feeds compiler errors and test failures back to the model. Only solutions that pass all verification tests are saved as synthetic training data.
2. Syntax-Aware SFT: The SLM is fine-tuned on this verified synthetic data to embed strong structural and syntactic priors specific to the target language.
3. Reinforcement Learning with Verifiable Reward (RLVR): The model undergoes further training using Group Relative Policy Optimization (GRPO), grounded by language-agnostic input/output tests. The SFT-trained model serves as a prior, constraining the search space and preventing the model from generating trivial syntax errors.

The pipeline significantly outperforms existing state-of-the-art methods on Julia and Ballerina, achieving substantial gains in pass@1 rates on benchmarks like MultiPL-E and Agnostics LiveCodeBench. Notably, the approach achieves these results using only one-third of the data and one-sixth of the computational cost of previous methods. The authors demonstrate that the pipeline effectively bootstraps code generation for languages with near-zero pretraining representation, such as Ballerina, by relying entirely on compiler and test-case feedback.

This research provides a scalable, cost-effective recipe for adapting LLMs to new or niche programming languages. By shifting the burden of iterative refinement from inference time to a one-time offline data curation process, the authors enable the deployment of high-performance, efficient models that do not require expensive online compute resources.