Medicine

Compositional Reasoning Depth Predicts Clinical AI Failure: Empirical Evidence Consistent with Transformer Compositionality Limits in Electronic Health Record Question Answering

Sanjay Basu

Featured June 21, 2026

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

Big AI models struggle to answer complex medical questions from patient records if they need to connect many pieces of information, and even asking them to "think step-by-step" doesn't fix this multi-step reasoning problem.

In depth

The paper reveals that large language models struggle systematically with multi-step reasoning in clinical settings. It introduces a "hop-count" metric to quantify the inferential steps needed for an EHR question, showing that model accuracy consistently drops as this hop count increases. Crucially, even advanced techniques like chain-of-thought prompting or retrieval augmentation fail to overcome this fundamental limitation, highlighting a critical safety concern for clinical AI deployment.

Key Takeaways

1
A hop-count taxonomy quantifies the compositional reasoning depth of clinical EHR questions, from single-fact retrieval (hop=1) to multi-component synthesis (hop=4).
2
Large language models exhibit a monotone accuracy decline as the hop count of clinical EHR questions increases, a pattern replicated across multiple models and providers.
3
Neither extended thinking (Chain-of-Thought) nor retrieval-augmented generation (RAG) significantly flattens the observed accuracy-versus-depth curve, indicating a persistent limitation in multi-hop clinical reasoning.

Conceptual Flow

HIGH LEVEL

Methodology (The "Logic")

The study measured how well AI models answered medical questions by counting how many thinking steps each question needed, from simple facts to complex summaries.

Medical Question

Patient Record

Count Thinking Steps

Question Difficulty Level

Results (The "Impact")

They found that the more thinking steps a question needed, the worse the AI models performed, showing a clear limit to their ability to reason deeply.

Question Difficulty Level

AI Model Answer

Measure Correctness

Accuracy Drops with Difficulty

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.