NLP has evolved from rule-based systems to Transformer-based LLMs, revolutionizing human–machine interaction. However, modern NLP systems face key bottlenecks: data privacy concerns in centralized training, limited adaptability to dynamic environments, and weak grounding in real-world intelligent systems. Recent advances combining FL, NLP, and RL have enabled privacy-preserving, communication-efficient, and aligned intelligence. This review contextualizes these developments through benchmark settings, datasets, models, and communication metrics. In NLP, LoRA-based federated fine-tuning reduces trainable parameters and transmitted data by 30–75%, achieving competitive accuracy on GLUE, ATIS, SNIPS, and clinical NER tasks (i2b2 and MIMIC-III) while maintaining privacy across distributed clients. Compared to full-model updates, parameter-efficient methods yield 100–1000× communication reduction, further improved by quantization (4–8×), gradient sparsification (10–100×), and layer skipping (∼70%). Under heterogeneous device conditions (up to 1000× compute variation), FedAvg and FedProx reach within 1–3% and 5–10% of centralized accuracy, while ILoRA improves convergence stability by 4–8%. In parallel, NLRL enhances sample efficiency by 15–25% through language-guided policy optimization, and RLHF-based alignment boosts human preference metrics by 10–30%. These techniques collectively enable scalable, secure, and adaptive learning across distributed and embodied environments. Finally, this survey highlights the synergy among FL, RL, and NLP, showing their integration within intelligent edge and robotic systems, achieving over 95% accuracy under constrained resources, and outlines open challenges in bias mitigation, communication bottlenecks, and safety in autonomous decision-making.

- First author: Tesfahunegn Minwuyelet Mengistu (Department of Information Convergence Engineering, Pusan National University)
  - Corresponding author: Taewoon Kim (Department of Information Convergence Engineering, Pusan National University)
  - Title of original paper: Natural language processing at the crossroads: Integrating federated and reinforcement learning for emerging intelligent systems
  - Journal: Computer Science Review
  - Web link: https://doi.org/10.1016/j.cosrev.2026.101014
  - Contact e-mail: taewoon@pusan.ac.kr