Large Language Models (LLMs) are rapidly advancing across diverse domains, yet their application in theoretical physics remains inadequate. While current models show competence in mathematical reasoning and code generation, we identify critical gaps in physical intuition, constraint satisfaction, and reliable reasoning that cannot be addressed through prompting alone. Physics demands approximation judgment, symmetry exploitation, and physical grounding that require AI agents specifically trained on physics reasoning patterns and equipped with physics-aware verification tools. We envision physics-specialized AI agents that handle multimodal data, propose physically consistent hypotheses, and autonomously verify theoretical results, and we call for collaborative efforts between physics and AI communities to build the specialized infrastructure necessary for AI-driven scientific discovery.

Exact scientific discovery requires more than heuristic search: candidate constructions must be turned into exact objects and checked independently. We extend TeXRA with an independent Lean 4 verification layer, turning it into a human-guided multi-agent platform that couples symbolic synthesis, combinatorial and linear-programming search, exact reconstruction of numerical candidates, and formal verification in Lean. Applying it to nonadditive quantum error-correcting codes with prescribed transversal diagonal gates, we obtain a Lean-certified catalogue of 14,116 distance-2 codes and resolve the transversal-T problem for several distance-3 codes, with all constructions, infinite families, and no-go results formalized and checked in Lean.

Adversarial machine learning studies vulnerabilities of machine-learning models in adversarial settings and develops techniques to make learning robust. We show that quantum classifiers, like their classical counterparts, are vulnerable to adversarial examples: adding carefully crafted, imperceptible perturbations to legitimate inputs leads to misclassification, demonstrated for classifying real-life images, phases of matter, and quantum data. We also show that practical defense strategies can be designed to counter a range of such attacks, bridging machine learning and quantum physics and offering guidance for implementing quantum classifiers on near-term and future quantum devices.