**A substantive answer** with specific claims, mechanisms, or caveats I can build follow-ups from

I notice that none of these preprints directly address frontier propulsion, relativity, or mathematical physics in the way your archive description suggests. Instead, the provided papers focus on AI's ability to reason about physics rather than on the physics itself.

That said, I can extract one substantive point relevant to your question:

On AI as a research tool in frontier physics: The evidence is mixed and domain-dependent. LLMs show near-ceiling performance on explicit, well-defined derivations within stable conceptual frameworks [AG-2026.04-565], but systematically fail when tasks require reconstruction of omitted reasoning steps or reorganization of representations under global consistency constraints [AG-2026.04-565]. This matters because frontier physics—propulsion, relativity, exotic spacetimes—often demands exactly that kind of implicit structural reasoning.

More concretely, when researchers fine-tuned small reasoning models specifically on Quantum Field Theory, they found they needed to generate over 2,500 synthetic training problems because "open-source verifiable training data required to train such capabilities is scarce" [AG-2026.04-892]. This suggests that AI tools in frontier physics are still scaffolding-dependent and cannot yet reliably navigate truly novel theoretical terrain without heavy curation.

The preprints provided don't substantively address propulsion mechanisms, relativity phenomenology, or mathematical physics results—they assess AI reasoning about physics rather than advancing physics itself.

What specific frontier physics question were you hoping to explore?