Agentic AI
New AI Framework Helps LLMs Trace Multiple Causes of Pollution and Disease
Researchers introduce IFAR, a zero-shot reasoning method that enables large language models to identify root causes across different levels and perspectives, boosting performance by 40%.
A research team has developed a new reasoning framework that significantly improves how large language models can work backwards from a problem to identify its underlying causes. The approach, called Inverse-Forward Abductive Reasoning (IFAR), tackles a longstanding weakness in AI reasoning by handling situations where multiple causes at different levels contribute to outcomes like pollution or disease spread.
The Challenge of Multi-Level Reasoning#
Current LLMs struggle with abductive reasoning—the ability to infer likely causes from observed effects. The real-world difficulty lies in recognizing that causes exist on multiple levels and from multiple angles. A disease outbreak, for instance, might stem from environmental factors, human behavior, infrastructure gaps, and policy failures all at once. No existing method had adequately addressed this layered complexity.
How IFAR Works#
The IFAR framework operates in a two-stage process: it begins with generalized backward reasoning to hypothesize potential causes, then validates each lead through relation-by-relation forward verification. Importantly, the method requires no special training—it works as a zero-shot approach compatible with standard LLMs. The researchers also created a specialized dataset called DeepAbduction focused on tracing causes in pollution and disease scenarios, filling a gap in available benchmarks for this type of reasoning task.
Performance Gains#
Experimental results show IFAR delivers roughly a 40% improvement in F1 score compared to alternative methods across major LLM architectures. The framework maintains a strong balance between catching all relevant causes (recall) and avoiding false leads (precision). Notably, IFAR lifted the reasoning abilities of non-specialized LLMs above those of models explicitly trained for complex reasoning tasks—and continued to enhance already-capable models further.
The authors plan to release their code following publication of the work.
