Quantum Architecture Search (QAS) enables the automated generation of quantum gate circuits tailored to specific computational tasks, demonstrating exceptional efficacy in quantum state preparation, quantum image processing, and molecular ground-state energy calculation. Differentiable Quantum Architecture Search (DQAS) significantly enhances search efficiency through continuous relaxation and gradient descent methods. However, existing differentiable techniques struggle to effectively balance exploration and exploitation, leading to degraded precision and redundant computations. To address this, a Temperature-Feedback Adaptive Differentiable Quantum Architecture Search (TFA-DQAS) is proposed, which dynamically adjusts search strategies by real-time monitoring of convergence trends. Experimental results indicate that this method achieves ground-state energy prediction accuracies of \(4.3\times 10^{-10}\) (H \(_2\) ), \(2.0\times 10^{-5}\) (LiH), \(1.1\times 10^{-6}\) (TFIM), and \(2.9\times 10^{-5}\) (Heisenberg), with peak precision surpassing DQAS by 1.3–10 \(^4\) -fold. Overall performance also outperforms existing approaches. The adaptive nature of TFA-DQAS renders it particularly suitable for near-term quantum devices constrained by circuit depth and gate count.