Three new papers on arXiv address different aspects of large language model development and deployment.
OpenTinker introduces an infrastructure for reinforcement learning of LLM agents, according to arXiv paper 2601.07376v1. The system is built around “a separation of concerns across algorithm design, execution, and agent-environment interaction,” departing from monolithic approaches.
Semantic Information Theory for LLMs is explored in arXiv paper 2511.01202v2, titled “Forget BIT, It is All about TOKEN.” The paper notes that while LLMs have “demonstrated remarkable capabilities in numerous real-world applications,” experimental research is “progressing rapidly” but “demands substantial computation.”
ARCQuant addresses LLM quantization in arXiv paper 2601.07475v1. The research focuses on fine-grained numerical formats like NVFP4, which the authors say “presents new opportunities for efficient Large Language Model (LLM) inference.” However, the paper notes “it is difficult to adapt existing Post-Training Quantization (PTQ) strategies to these formats.” ARCQuant proposes using “Augmented Residual Channels” to improve NVFP4 quantization.
All three papers represent ongoing efforts to make LLMs more efficient and effective across different dimensions: training infrastructure, theoretical foundations, and deployment optimization.