CRAFT Enables Explainable AI for 6G RAN Networks

Researchers have introduced CRAFT (Cold-start Reasoning Alignment via Fine-Tuning), a novel data-centric method aimed at achieving pre-hoc explainability in AI-native 6G Radio Access Networks (RAN). The vision for 6G networks includes embedding Small Language Models (SLMs) for real-time reasoning, but existing training paradigms for telecom LLMs primarily offer post-hoc rationalization, making decision processes unauditable. CRAFT addresses a "cold-start barrier" encountered when trying to transplant reinforcement learning methods like Group Relative Policy Optimization (GRPO) into telecom settings, where SLMs struggle to learn both output format and label prediction simultaneously. CRAFT autonomously generates a verified dataset of (input, trace, label) triplets, then fine-tunes SLMs using LoRA, requiring substantially less compute and time than GRPO. It achieves high accuracy and F1 scores (up to 86.5% and 94.6%) on telecom datasets with no parse failures, consumes 59% less energy, and provides a robust foundation for further GRPO fine-tuning, paving the way for auditable AI in 6G RAN.