BEAVER: A Multi-Agent Framework for Evidence-Grounded Answering and Property-Constrained Design of Degradable Polymers

<jats:p>Degradable polymers are an important research area in materials science because of their environmental compatibility and tunable degradation. Scientific agents could accelerate discovery by integrating fragmented literature with heterogeneous property data, but two challenges remain: retrieving relevant evidence with sufficient coverage and converting it into traceable reasoning and multi-property-constrained material design. Here, we present BEAVER, an evidence-grounded multi-agent framework for scientific question answering and degradable polymer design. Its Intent-Planner architecture coordinates hybrid retrieval-augmented generation in two workflows: Query Mode for evidencegrounded question answering and Design Mode for candidate generation and assessment under property constraints. Across retrieval, question-answering, and design benchmarks, BEAVER translated higher-quality evidence into more reliable outputs. Reranking increased the overall RAGAS score by more than 0.08 on average across retrieval settings, while the reranked chunk-level hybrid configuration achieved the best balance between recall and relevance. In question answering, BEAVER obtained a bidirectional mean Overall score of 8.76, compared with 8.33 for the strongest closed-book baseline. In coupled-property design, it performed comparably to representative frontier LLMs while providing clearer evidence traceability, more complete constraint handling, and more 1 explicit analysis of risks and uncertainties. Overall, BEAVER offers a traceable and extensible framework that unifies hybrid retrieval, evidence-grounded reasoning, and property-constrained design for degradable polymer research.</jats:p>