Exploring the Viability of Automated Heuristic Design for 5G LDPC Decoding

Abstract

Automated Heuristic Design (AHD) leverages Large Language Models (LLMs) and task-specific evaluation to search among existing algorithmic components and generate novel ones. This paper studies AHD for 5G New Radio (NR) Low-Density Parity Check (LDPC) decoding by evolving the Check Node Update (CNU) function used in iterative Belief Propagation (BP). To this end, we implement a flexible AHD framework capable of accommodating different evolution policies and assess their performance (i.e., the decoding accuracy achieved by the discovered heuristic) and computational complexity (i.e., the number of evaluated candidate heuristics) in the target 5G NR task. We experiment with two evolution policies: LLM- Based Evolution (LBE), which performs population-based par- allel mutation and selection, and Prompt-guided LLM-Based Evolution (PLBE), which augments evolution with structured prompt operators. Under a fixed time budget per experiment, we find that AHD prompted by specifying the context of the task (i.e., LDPC decoding) consistently converges toward state- of-the-art performance and is robust to the specific evaluation approach employed. Instead, context-agnostic prompting and/or exploratory parent sampling tend to stagnate at substantially lower scores. The best discovered CNU heuristic is structurally close to functions employed in production 5G networks and marginally outperforms such functions on the specific Transport Block (TB) batch used for AHD. However, the additional gain disappears under independently drawn TBs. Ultimately, our study highlights the promise of AHD for 5G tasks but also the need for careful validation when interpreting in-loop gains.