Four research papers published on arXiv examine critical challenges in large language model development and deployment.
According to arxiv.org, a comprehensive survey on federated prompt learning (FPL) reviews recent advances in integrating federated learning with LLMs. The paper addresses how FPL enables “clients to collaboratively train a learning model without sharing raw data, making it a promising solution for privacy-preserving LLM training and reasoning.” The survey examines FPL methods across the full model lifecycle, including pre-training, fine-tuning, and practical applications.
Separately, arxiv.org reports on P2Skill, a privacy-preserving method for cloud-local LLM inference systems. According to the paper, P2Skill uses “a prompt-based skill distillation method in which a local small language model (SLM) autonomously performs decomposition, PII-aware routing, paraphrasing, and reconstruction.” The evaluation shows P2Skill achieves “1.69× and 3.66× higher privacy-preserved inference quality than previous baselines” on a four-domain benchmark.
According to another arxiv.org paper published in COLM 2026, researchers examined reasoning behaviors in thinking models across 15 models and 6 benchmarks. The study annotated 15,282 traces and found evidence for an “Amplification-Lift Gap,” where thinking models “strongly amplify self-correction, hypothesis testing, and uncertainty acknowledgment,” while “confidence calibration, knowledge alignment, and self-awareness” showed the highest association with correct answers.
Additionally, arxiv.org reports on BAT, a system accepted to ICML 2024 that combines spatial sound perception with LLM reasoning capabilities to interpret spatial audio environments.