Grounding LLMs in Scientific Discovery via Embodied Actions

Large Language Models (LLMs) have shown significant potential in scientific discovery but struggle to bridge the gap between theoretical reasoning and verifiable physical simulation. Existing solutions operate in a passive "execute-then-response" loop and thus lack runtime perception, obscuring agents to transient anomalies (e.g., numerical instability or diverging oscillations). To address this limitation, we propose EmbodiedAct, a framework that transforms established scientific software into active embodied agents by grounding LLMs in embodied actions with a tight perception-execution loop. We instantiate EmbodiedAct within MATLAB and evaluate it on complex engineering design and scientific modeling tasks. Extensive experiments show that EmbodiedAct significantly outperforms existing baselines, achieving SOTA performance by ensuring satisfactory reliability and stability in long-horizon simulations and enhanced accuracy in scientific modeling.

When scientists use AI assistants to run scientific simulations, the AI usually writes the code, hits run, and waits for the final answer. The trouble is that simulations can go quietly wrong in the middle, with numbers drifting or oscillations spiraling out of control. The AI has no way of noticing until everything finishes and the result is already wasted. We built a system called EmbodiedAct that lets an AI watch its simulation as it runs, the way a pilot watches the instruments rather than only reading the flight log afterward. The AI stays connected to the simulation software (we used MATLAB and Simulink) the whole time. It sees what is happening step by step and can step in to fix things when the simulation starts going off course. We tested EmbodiedAct on engineering design problems and physics modeling tasks. It produced more reliable and accurate results than existing AI methods that only check the outcome at the end. The work points toward AI assistants that can carry out scientific experiments in realistic simulated environments, not just write code and hope for the best. It brings autonomous scientific discovery a step closer to something human researchers could actually trust.