llm-simulator

DSWorld introduces a Data Science World Model that predicts environment state transitions to reduce costly trial-and-error in autonomous agents, achieving 14x acceleration in RL training and 3-6x in inference while maintaining competitive performance.

Presents Onnes, a physics-grounded multi-agent LLM simulator for cryogenic fault diagnosis in quantum computing infrastructure. With curated few-shot demonstrations and self-consistency voting, a zero-shot LLM agent panel achieves 0.990 fault-classification accuracy, matching a supervised classifier without parameter updates.