Go to ICLR 2026 Conference homepageWhen Knowledge Hurts: Enriching Domain Knowledge for Causal Scientific Reasoning
Keywords: large language models, retrieval augmented generation, contextual tunneling, causal reasoning, material discovery
TL;DR: We solve the "tunneling problem" where LLMs+knowledge graphs hurt performance by over-focusing on narrow contexts—ARIA uses causal reasoning to recover lost performance and achieve SOTA in scientific discovery.
Abstract: Science has long sought to uncover principles of discovery, yet fields like materials science remain slow and labor-intensive. While Large Language Models (LLMs) can accelerate progress by integrating domain knowledge, we reveal a critical failure mode: \textbf{\textit{contextual tunneling}}, where naive knowledge integration causes LLMs to over-anchor on narrow retrieval paths while suppressing broader parametric reasoning. Through the evaluation in materials discovery, we demonstrate that naive knowledge graph augmentation degrades performance by 15–35\% on key reasoning tasks compared to direct prompting.
To address this challenge, we introduce \texttt{ARIA} (Autonomous Reasoning Intelligence for Atomics), a causal-aware framework featuring: (i) hierarchical reasoning that provides graceful degradation to knowledge graph sparsity, (ii) enhanced analogical transfer for robust reasoning, (iii) knowledge graph enrichment through online searching. Extensive experiments show that while naive KG integration consistently underperforms baseline LLMs, \texttt{ARIA} not only recovers this loss but also provides interpretable causal explanations by tracing reasoning through the knowledge graph, enabling scientists to verify and trust its outputs. Our work demonstrates that external knowledge can inadvertently constrain reasoning and establishes a principled framework for robust KG–LLM integration in scientific discovery.
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
Submission Number: 22734
Loading