Agentic AI
AI system with built-in safety checks tackles quantum chemistry simulations reliably
Researchers introduce DREAMS, a multi-agent framework that combines large language models with rigorous verification to make autonomous materials simulation trustworthy enough for scientific use.
Large language models can orchestrate complex scientific workflows over many steps, but their outputs often raise a fundamental problem: outputs may seem plausible while being fundamentally wrong, and the systems can lose track of their reasoning chains. A new framework called DREAMS tackles this challenge head-on by weaving verification checks throughout an AI agent's decision-making process.
A Guard Rail Built into Scientific Reasoning#
The system, developed around density functional theory (DFT) simulations for materials science, employs a multi-tier safety architecture. Rather than letting an LLM agent roam freely, DREAMS validates every decision: some checks are deterministic (applying hard rules where they exist), while others delegate judgment to the AI model itself—but always one parameter at a time. Critically, every value is traced back to its registered source, preventing the agent from inventing or laundering numbers. This verification happens at two key moments: when tools are first called (blocking fabricated or unsourced inputs before they contaminate the workflow), and when results are compiled into reports (auditing the entire chain of reasoning behind each claim).
Benchmarks Show Near-Expert Accuracy#
Testing on materials simulation tasks reveals the payoff. On the Sol27LC lattice-constant benchmark, the guarded system achieved average errors below 1%, on par with expert work. For a tricky chemistry puzzle—predicting how carbon monoxide adsorbs onto platinum surfaces—DREAMS reproduced expert-level results and used Bayesian ensemble sampling to map uncertainty driven by the choice of quantum functional. It even confirmed that a face-centered-cubic site is energetically favored under the generalized gradient approximation level of theory.
The Cost of Trustworthiness#
There is a trade-off. The unguarded version of the same system reached nearly correct answers while only 81% of its essential steps succeeded; the fully guarded version uses about 13 times more input tokens to verify every step. However, this design flexibility allows users to dial verification up or down individually, balancing rigor against computational cost. When researchers tested their validation rules across five different judge models, the rules transferred well, suggesting the approach may generalize beyond a single AI variant.
