Comet-H: Orchestrating LLMs for Research Software with Evolving Specs

A new arXiv preprint (2604.27209) identifies two failure modes in using large language models for research-software projects: hallucination accumulation, where claims outpace code or theory, and desynchronization, where code, theory, and the model's world model fall out of alignment. The authors propose Comet-H, an iterative prompt automaton that coordinates ideation, implementation, evaluation, grounding, and paper-writing as coupled components of a single workspace state. A controller selects prompts based on workspace gaps and carries forward unfinished work with a half-life mechanism.