Recent large language models (LLMs) have made weight-space programming (WSP) feasible, allowing neural network weights to be generated directly without gradient‑based optimization. Extending WSP to physical problems faces two challenges: achieving numerical accuracy while respecting physical conservation or equation constraints. To address this, we introduce the Physics2Weight (PtW) framework, which generates networks for partial differential equations (PDEs) through three mechanisms:\n\n- Physics‑constrained prompting: embed the governing equation, initial condition and boundary condition symbols into the LLM prompt so that the produced weights are inherently physics‑aware.\n- Physics‑guided architecture editing: automatically add or remove layers and adjust activation functions based on operator characteristics of the equation (e.g., second‑order diffusion, convection terms).\n- Iterative weight refinement: after the initial generation, use the self‑regressive capability of WSP to refine weights over multiple rounds, progressively reducing residuals without any gradient steps.\n\nWe evaluate PtW on four PDEs: the heat equation $u_t = \alpha \nabla^2 u$, the advection equation $u_t + \mathbf{c}\cdot\nabla u = 0$, Burgers' equation $u_t + u\nabla u = \nu \nabla^2 u$, and the Allen‑Cahn equation $u_t = \epsilon^2 \nabla^2 u - f(u)$. Under identical mesh and training budget, PtW consistently yields lower solution errors than conventional physics‑informed neural networks (PINNs), especially for the linear heat and advection cases. However, for the nonlinear Burgers and Allen‑Cahn equations, the generated weights still struggle to fully satisfy the residual constraints, indicating that current WSP methods need further improvement for strongly nonlinear regimes.\n\nOverall, PtW demonstrates the potential of LLMs to construct physics‑consistent neural networks without gradients, offering a new direction for accelerating scientific computation.\n\nReview