Bootstrapping Code Translation with Weighted Multilanguage Exploration

TL;DR

BootTrans tackles the scarcity of multilingual parallel code with executable test oracles and the optimization imbalance across language directions by bootstrapping from a pivot language's unit tests and expanding training via execution-verified rollouts. It introduces a dual-pool RL framework and a language-aware weighting scheme (GRPO) to progressively build a multilingual translation policy that improves functional correctness across all language pairs. Empirical results on HumanEval-X and TransCoder-Test show substantial gains over strong baselines across six translation directions and multiple model families, with ablations validating the contributions of both bootstrapping and weighting. The approach offers a scalable pathway to robust multilingual code translation and demonstrates compatibility with inference-time enhancement methods, highlighting its practical potential for cross-language software modernization.

Abstract

Code translation across multiple programming languages is essential yet challenging due to two vital obstacles: scarcity of parallel data paired with executable test oracles, and optimization imbalance when handling diverse language pairs. We propose BootTrans, a bootstrapping method that resolves both obstacles. Its key idea is to leverage the functional invariance and cross-lingual portability of test suites, adapting abundant pivot-language unit tests to serve as universal verification oracles for multilingual RL training. Our method introduces a dual-pool architecture with seed and exploration pools to progressively expand training data via execution-guided experience collection. Furthermore, we design a language-aware weighting mechanism that dynamically prioritizes harder translation directions based on relative performance across sibling languages, mitigating optimization imbalance. Extensive experiments on the HumanEval-X and TransCoder-Test benchmarks demonstrate substantial improvements over baseline LLMs across all translation directions, with ablations validating the effectiveness of both bootstrapping and weighting components.

Figures (5)

• Figure 1: Illustration of challenges in scaling RLVR for multilingual code translation. (i) Input monotony: verifiable seeds are predominantly confined to a single pivot language. (ii) Optimization imbalance: varying task difficulties lead to biased learning signals.
• Figure 2: Overview of BootTrans. It comprises two key components: (i) Bootstrapping Multilanguage Exploration (left and center panels), which expands the training set via execution-verified translations; (ii) Language-aware Weight Optimization (right panels), which dynamically re-weights the loss via cross-lingual performance gaps.
• Figure 3: Performance comparison of different pivot languages with Qwen3-1.7B.
• Figure 4: Two examples of C++ translated into Python. BootTrans succeeds in the top case but fails in the bottom.
• Figure 5: Error classification on HumanEval-X.