How much does fine-tuning on one physics domain actively harm transfer to structurally different fields?

Your question asks about negative transfer—whether specializing a model on one physics domain actively degrades its performance elsewhere. This is a genuine concern in machine learning, but the provided preprints don't directly measure the harm.

Here's what the archive *does* show:

Positive transfer is achievable across physics domains. When researchers fine-tuned a particle-flow reconstruction model trained on one collider detector (CLICdet) and transferred it to a structurally different detector (CLD), they achieved the same performance with 10× fewer samples from scratch [AG-2025.02-1472]. Similarly, generative shower simulation models using point clouds transferred across detector geometries with only 100 target-domain samples, yielding 44% improvement over training from scratch [AG-2025.11-1571].

But specialized fine-tuning in one physics domain shows mixed generalization. When researchers fine-tuned a 7B reasoning model specifically on Quantum Field Theory (QFT) problems, they benchmarked "generalization to other physics domains," but the abstract doesn't quantify whether QFT specialization *harmed* performance elsewhere [AG-2026.04-892]. A separate study found that LLMs near ceiling on explicit QFT/string theory derivations but systematically degrade when reasoning steps are omitted or representations need reorganization [AG-2026.04-565]—suggesting domain-specific brittleness, though not negative transfer per se.

The honest answer: The archive doesn't measure how much QFT fine-tuning actively *harms* cosmology reasoning, or vice versa. You'd need a direct comparison of (1) base model → domain A → measure domain B performance versus (2) base model → measure domain B performance directly. That experiment appears absent here.