On-Policy Optimization with Group Equivalent Preference for Multi-Programming Language Understanding

TL;DR

This paper introduces a novel reinforcement learning framework called OORL with Group Equivalent Preference Optimization (GEPO) to enhance large language models' ability to understand and translate code across multiple programming languages, improving their functional comprehension.

Contribution

The paper proposes a new RL framework combining on-policy and off-policy strategies with a preference optimization method based on IR groups, advancing multi-language code understanding in LLMs.

Findings

Significant performance improvements on multi-language code benchmarks.

Effective transfer of coding proficiency across diverse programming languages.

Enhanced recognition of code functionality and relationships between different language implementations.

Abstract

Large language models (LLMs) achieve remarkable performance in code generation tasks. However, a significant performance disparity persists between popular programming languages (e.g., Python, C++) and others. To address this capability gap, we leverage the code translation task to train LLMs, thereby facilitating the transfer of coding proficiency across diverse programming languages. Moreover, we introduce OORL for training, a novel reinforcement learning (RL) framework that integrates on-policy and off-policy strategies. Within OORL, on-policy RL is applied during code translation, guided by a rule-based reward signal derived from unit tests. Complementing this coarse-grained rule-based reward, we propose Group Equivalent Preference Optimization (GEPO), a novel preference optimization method. Specifically, GEPO trains the LLM using intermediate representations (IRs) groups. LLMs can be guided to discern IRs equivalent to the source code from inequivalent ones, while also utilizing signals about the mutual equivalence between IRs within the group. This process allows LLMs to capture nuanced aspects of code functionality. By employing OORL for training with code translation tasks, LLMs improve their recognition of code functionality and their understanding of the relationships between code implemented in different languages. Extensive experiments demonstrate that our OORL for LLMs training with code translation tasks achieves significant performance improvements on code benchmarks across multiple programming languages.