How do human physicists acquire this tacit structural knowledge during training versus how LLMs learn from text?

This is a question about cognitive science and machine learning pedagogy rather than frontier physics — and it's an excellent one, but it sits at the boundary of what this archive covers.

That said, the preprints here offer some empirical windows into the gap you're asking about.

What human physicists do:

Human training in theoretical physics involves *embodied practice*. A graduate student doesn't just read QFT; they work problem sets for years, internalize why certain Feynman diagrams contribute and others vanish, feel the constraints of gauge invariance in their bones. They develop what Michael Polanyi called "tacit knowledge" — understanding that lives below the level of explicit statement.

What LLMs do (and don't):

The archive suggests LLMs acquire something different. When researchers evaluated contemporary LLMs on QFT and string theory using a five-level rubric, they found a striking pattern: models perform "near-ceiling" on explicit derivations within stable conceptual frames, but "systematically degrade when tasks require reconstruction of omitted reasoning steps or reorganization of representations under global consistency constraints" [AG-2026.04-565]. In other words, LLMs excel at pattern completion within established scaffolds but struggle with the *unspoken structural knowledge* that human physicists internalize.

When researchers fine-tuned smaller models (7B parameters) on 2,500+ synthetic QFT problems, they observed that "reasoning errors evolve during RL and SFT," and they could track this in the models' chains-of-thought [AG-2026.04-892]. The implication: LLMs *can* develop something closer to tacit reasoning, but it emerges through specialized training on domain problems — not passively from general text.

This mirrors what happens in human training: repetition and struggle with ill-posed problems, not passive absorption of text.

The deeper issue:

When researchers reverse-engineered LLMs' representations of QCD concepts, they found "naturally idiosyncratic patterns" in how foundational principles embed in parameter space [AG-2025.11-1583]. This suggests LLMs develop *representations*, but not necessarily the ones a human physicist builds through years of solving problems and getting corrected by advisors and peers.

A human physicist's tacit knowledge is embodied in intuition honed through failure; an LLM's emerges from statistical patterns in token sequences. The gap isn't just one of scale — it's about the *ontology* of how understanding forms.