Beyond Scaling Laws: Towards Scientific Reasoning-Driven LLM Architectures

Large language models (LLMs) are transforming the practice of scientific research, with applications ranging from literature synthesis and hypothesis generation to molecular design and experimental planning. Yet despite their linguistic fluency and scale, current LLMs struggle with scientific reasoning tasks that require structured logic, physical constraints, causal inference, and symbolic manipulation. In this Perspective, we argue that these limitations stem not from insufficient data or compute, but from architectural misalignment with the epistemic demands of science. We propose that next-generation scientific LLMs must move beyond token-level prediction to embrace structure-augmented architectures. In particular, we highlight two key design principles: the integration of graph neural networks (GNNs) to capture relational scientific structures, and the deployment of modular multi-agent systems that reflect the distributed, iterative nature of scientific inquiry. Together, these innovations can transform LLMs into reasoning engines capable of hypothesis testing, simulation coordination, and collaborative discovery. We further call for open, domain-specific, and interpretable LLM ecosystems as foundational infrastructure for the next paradigm of science.