Skill-Contracted Agents for Evidence-Aware Materials Literature Analysis

Short summary

AlphaAgent is a skill-driven agent framework for materials science literature analysis that decouples retrieval-based QA from report generation through explicit skill contracts. It queries 300K+ metallurgy papers, reformulates insufficient queries, and produces structured per-paper and cross-paper reports. In blind evaluation on 40 questions, it substantially outperformed a matched baseline, with largest gains in mechanistic explanation and credibility awareness.

• •AlphaAgent decouples retrieval QA from report generation via skill contracts for materials science
• •Queries 300K+ metallurgy papers with query reformulation when evidence is insufficient
• •Outperforms matched baseline on 40 blind questions, especially in mechanistic reasoning

Generated with AI, which can make mistakes.