AI
New Framework Combines AI with Simulations to Fix Hallucinations in Scientific Q&A
Researchers propose SimulRAG, a retrieval system that grounds language models in scientific simulators to improve accuracy and trustworthiness when answering complex scientific questions.
Large language models excel at synthesizing complex information and crafting persuasive explanations, but they frequently invent details or make unsupported claims when answering scientific questions at length. SimulRAG, a new approach developed by researchers and documented in an arXiv paper, tackles this by anchoring AI-generated answers in evidence from scientific simulators.
The Core Problem#
When LLMs tackle long-form scientific questions, they often produce hallucinations—claims that sound plausible but lack grounding in actual evidence. Traditional retrieval-augmented generation (RAG) improves matters by pulling in documents, but it cannot validate numerical predictions or trace how systems evolve over time. Scientific simulators offer something documents cannot: the ability to test hypotheses quantitatively and model dynamic processes. Integrating these simulators into the retrieval pipeline, however, presents two major obstacles: how to query a simulator using natural language, and how to efficiently verify and refine lengthy, multi-claim answers.
How SimulRAG Works#
The framework uses a generalized retrieval interface that converts between human-readable text and simulator parameters or outputs, bridging the gap between natural language and simulation code. Rather than revising entire answers at once, SimulRAG operates at the granular level of individual claims, pairing each with an uncertainty score and a judgment about whether the simulator can check it.
The method first asks the model to generate diverse candidate answers without any retrieval, then selectively refines uncertain claims that a simulator can validate. This targeted approach avoids unnecessary computation while homing in on the most error-prone statements.
Validation and Results#
The researchers released a benchmark for evaluating long-form scientific QA spanning three domains: climate science, epidemiology, and urban planning. Ground truth answers were verified by running simulations and obtaining human review. Testing SimulRAG against adapted versions of existing RAG systems, the framework boosted informativeness by 30.4% and factuality by 16.3%, demonstrating substantial gains in both completeness and accuracy.
