Learning Relation-Agnostic Knowledge Graph Traversal for Materials Science Question Answering

Materials science question answering requires connecting composition, processing, structure, and property, and has been approached with domain-specialized language models and tool-using agents. Knowledge graphs (KGs) offer a complementary route that keeps those connections explicit, but large-scale materials KGs are constructed from the literature by automated extraction, and what they encode is often coarse relation structure rather than curated semantics: in MatKG, a representative materials science KG, every one of the 49 relation labels is fully determined by the types of its endpoint entities. Existing KG-augmented agents guide traversal by relation semantics, which on such graphs carry no discriminative signal; we show that representative graph-reasoning methods then fail to outperform their zero-shot LLM baselines. We propose $\textbf{RAnGT}$, a $\textbf{R}$elation-$\textbf{A}$g$\textbf{n}$ostic Knowledge $\textbf{G}$raph $\textbf{T}$raversal agent that replaces relation-guided search with a traversal policy trained by group relative policy optimization (GRPO) against answer correctness and evidence quality, and a frozen Verifier that decides at inference whether retrieved evidence may influence the Answer Model. On the MaScQA materials science benchmark, RAnGT is the best method on all four open-weight backbones, improving over zero-shot answering by up to $+10.5$ points and over RoG, fine-tuned under an identical training budget, by $+5.0$ to $+18.7$ points, with the gain confined by construction to questions whose evidence the Verifier accepted. The policy, trained on only 65 questions, transfers unchanged to eight chemistry benchmarks and attains the best average accuracy among all compared methods, suggesting that verification-gated, learned retrieval turns noisy materials science KGs from a liability into a usable evidence source.