FunPRM Enhances Large Language Model Code Generation Through Modular Prompting and Reward Correction

Global: FunPRM Enhances Large Language Model Code Generation Through Modular Prompting and Reward Correction

Researchers have introduced FunPRM, a test‑time scaling technique that improves the ability of large language models (LLMs) to generate complex code. The method was evaluated in 2026 on two public benchmarks, LiveCodeBench and BigCodeBench, and demonstrated superior results across five base LLMs, including a new state‑of‑the‑art score on LiveCodeBench when paired with O4‑mini. By prompting LLMs to produce modular functions and applying a meta‑learning reward correction, FunPRM addresses both the lack of meaningful step decomposition and the noise inherent in Monte‑Carlo‑estimated partial‑solution rewards.

Background

Code generation remains a core application of LLMs, yet existing models frequently falter on tasks that require multi‑step reasoning or intricate program structure. Prior test‑time scaling approaches, such as Process Reward Model (PRM)‑based Best‑of‑N selection, have shown promise in mathematical reasoning but have struggled to translate to programming contexts because code does not naturally decompose into discrete reasoning steps.

Method Overview

FunPRM tackles these challenges in two ways. First, it prompts the LLM to organize output into separate functions, treating each function as an individual PRM reasoning step. This modular prompting encourages clearer, more maintainable code. Second, FunPRM incorporates a meta‑learning reward correction mechanism that leverages clean final‑solution rewards obtained from a unit‑test‑based evaluation system. The corrected rewards replace noisy partial‑solution scores, providing a more reliable signal for Best‑of‑N selection.

Experimental Evaluation

The authors conducted experiments on LiveCodeBench and BigCodeBench, two widely used code‑generation benchmarks. Five base LLMs were tested, each with and without FunPRM applied. Performance was measured using standard pass‑rate metrics derived from unit‑test outcomes.

Results

Across all five models, FunPRM consistently outperformed existing test‑time scaling methods. Notably, when combined with the O4‑mini model, FunPRM achieved the highest reported pass rate on LiveCodeBench, establishing a new state‑of‑the‑art benchmark. The approach also produced code that reviewers described as more readable and reusable, suggesting practical benefits beyond raw accuracy.

Implications and Future Work

By demonstrating that modular prompting and reward correction can substantially improve LLM‑driven code generation, FunPRM opens avenues for further research into structured prompting and meta‑learning techniques. Future investigations may explore extending the method to other programming languages, integrating additional static analysis tools, or applying the framework to downstream software‑development workflows.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.