[Paper Review] Grounding LLMs in Scientific Discovery via Embodied Actions

Bo Zhang, Jinfeng Zhou+4 more|arXiv (Cornell University)|Feb 24, 2026

Machine Learning in Materials Science0 citations

TL;DR

EmbodiedAct grounds LLMs in embodied actions with a tight perception-execution loop in MATLAB, enabling continuous monitoring and autonomous correction for reliable long-horizon scientific simulations and modeling.

ABSTRACT

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 lacks 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.

Motivation & Objective

• Bridge the gap between theoretical reasoning and verifiable physical verification in process-oriented scientific tasks.
• Transform existing scientific software into active embodied agents that couple execution with continuous perception.
• Demonstrate that runtime perception and hot-fix guided re-planning improve reliability, stability, and accuracy in long-horizon simulations.
• Show generalization across MATLAB and open-source backends, and across multiple model families.
• Provide a reusable framework and protocol for tight perception-action loops in science-focused LLM applications.

Proposed method

• Propose EmbodiedAct, a closed-loop cognitive architecture that grounds intent into executable simulation primitives.
• Ground LLMs in four modules: Strategic Planner, Primitive Generator, Runtime Perception Engine, and Reflective Decision Maker.
• Use an Asynchronous State Synchronization Protocol to maintain real-time, streaming observations from the simulation environment.
• Translate high-level scientific intents into software-specific primitives (e.g., MATLAB solvers, Simulink graph operations) via topological reasoning.
• Enable live monitoring of simulation trajectories and intermediate states to detect anomalies and trigger autonomous hot-fix repairs.
• Incorporate a dual-loop control: a fast inner loop for perception-driven hot-fixes and a slow outer loop for re-planning guided by physics-informed feedback.

Figure 1 : Comparison of EmbodiedAct with existing paradigms. EmbodiedAct integrates executable simulation primitives and continuous runtime perception, endowing the agent with the capacity for embodied action within physical simulation environments.

Experimental results

Research questions

• RQ1 Can an embodied agent with real-time perception outperform traditional code-as-action LLM agents in process-oriented scientific tasks?
• RQ2 How does runtime perception influence reliability, stability, and accuracy in long-horizon simulations across engineering and physical sciences?
• RQ3 Do domain-specific primitives and topology-aware execution improve solution quality compared to generic code generation?
• RQ4 What is the impact of a perception-action loop on generalization across MATLAB and open-source simulation backends?
• RQ5 Is there a measurable benefit of the dual-loop architecture (inner hot-fix loop + outer reflective planner) for complex systems?

Key findings

• EmbodiedAct achieves state-of-the-art performance on EngDesign and SciBench-107 benchmarks, outperforming generative and code-based baselines across multiple model families.
• Active runtime perception enables real-time monitoring of intermediate states and triggering of hot-fixes, improving reliability and reducing failure modes in long-horizon simulations.
• Domain-specific primitives and topology-aware translation of intent into simulation actions yield higher accuracy, especially in numerically sensitive domains like Mathematics and Physics.
• The approach generalizes across backends (MATLAB and open-source) with sustained performance advantages, demonstrating robust cross-environment applicability.
• Open-source models gain substantial uplift when paired with EmbodiedAct, narrowing the gap with proprietary models and highlighting the framework’s scaffolded strength.
• Ablation studies confirm that runtime perception is the key driver of performance, with perception-disabled variants approaching CodeAct performance.

Figure 2 : Overview of EmbodiedAct, which bridges the LLM agent and simulation environment via the Asynchronous State Sync Protocol. EmbodiedAct orchestrates a fast inner loop driven by the Runtime Perception Engine to trigger immediate Hot-Fixes, and a slow outer loop driven by the Reflective Decis