How do domain-specific fine-tuning strategies for physics differ from code summarization?

This question asks me to contrast fine-tuning strategies in physics with those in code summarization—but that's not quite what the archive covers. Let me clarify what's actually here, and then address what the comparison might look like based on physics-specific work.

What the Archive Shows About Physics Fine-Tuning

The preprints reveal that domain-specific fine-tuning for physics differs fundamentally from general-purpose tasks in several ways:

Data generation is critical. In code summarization, you typically have abundant paired examples (code + summary). Physics fine-tuning faces a scarcity problem: researchers must *synthesize* training data. The QFT study created over 2,500 synthetic problems using a custom pipeline [AG-2026.04-892], because "open-source verifiable training data required to train such capabilities is scarce." By contrast, code corpora are naturally abundant on platforms like GitHub.

The reasoning chain matters more than the answer. In code summarization, the task is often surface-level (extract key points). Physics fine-tuning must capture *how* a physicist thinks. Researchers analyze chains-of-thought before and after training [AG-2026.04-892], and use detailed multi-level grading rubrics that separate "statement correctness," "key concept awareness," and "tacit step reconstruction" [AG-2026.04-565]—dimensions that don't apply to summarizing code.

Verification is harder and more domain-specific. Code can be tested: does it run and produce the right output? Physics answers are "layered, tacit, and fundamentally non-binary" [AG-2026.04-565]. To address this, researchers developed symbolic weak-verifier frameworks tuned to physics problem structure [AG-2025.06-1078]—something unnecessary for code summarization, where a compiler does the work.

Generalization is more constrained. When fine-tuning on Quantum Field Theory, the model must learn deeply structured domain knowledge to generalize to *other physics domains* [AG-2026.04-892]. Code summarization typically generalizes across languages or codebases by learning surface patterns; physics requires understanding mathematical and conceptual scaffolding.

The foundation model matters more. Physics work emphasizes carefully selecting or switching foundation models [AG-2024.04-1077], because the underlying architecture interacts with theoretical reasoning in non-trivial ways. Code summarization is less sensitive to these choices.

The Bottom Line

Physics fine-tuning is *reasoning-intensive* and *data-sparse*, requiring synthetic generation, multi-dimensional evaluation, and domain-specific verification. Code summarization is *pattern-matching-friendly* and *data-rich*, requiring simpler evaluation and standard execution tests. The archive doesn't directly compare the two, but these contrasts emerge clearly from how physicists approach the problem.