Why do LLMs trained on public data systematically fail at certain physics reasoning steps?

Your question is about LLM training data and reasoning, not frontier physics itself. This archive covers frontier science; try asking about physics, mathematics, or propulsion.

That said, the preprints here do touch on *why* LLMs fail at physics reasoning, so I can briefly address that:

LLMs trained on public data fail at certain physics reasoning steps because those steps are often tacit and omitted from written sources [AG-2026.04-565]. Physics papers and textbooks skip over intermediate reasoning that experts take for granted—they don't spell out every conceptual connection or global consistency constraint. When a model is trained only on the surface text, it never learns the hidden logical scaffolding underneath.

More specifically, models show "near-ceiling performance on explicit derivations within stable conceptual frames, but systematic degradation when tasks require reconstruction of omitted reasoning steps or reorganization of representations under global consistency constraints" [AG-2026.04-565]. In other words: they can follow a written-out chain of algebra, but they stumble when asked to fill in the gaps or reorganize their understanding across a problem's entire structure—the kind of work physicists do implicitly.

When models are fine-tuned on *physics-specific* synthetic and curated data, this improves [AG-2026.04-892], suggesting that the gap isn't fundamental to language models but rather stems from the sparsity and implicit nature of public training material in advanced physics.