A recent benchmark shows a frontier language model can correctly guess the structure of a compound from its MS/MS spectrum at a 52 % rate for analytes that appear in literature, but falls to 2 % for unseen compounds. The test uses 100 spectra split evenly between “in‑distribution” and “out‑of‑distribution” cases, matched on adduct, instrument, precursor m/z, and peak count (f2). (f5, f3). Adding subformula annotation or forward prediction alone only improves the 52 % baseline for in‑distribution cases; they do not change the out‑of‑distribution score (f3). The performance ceiling for out‑of‑distribution prediction equals the generator’s candidate coverage: if the true structure is not among the generated candidates, the model cannot rank it first, limiting accuracy to the generator’s coverage even though the model can re‑rank candidates efficiently (f4). These results suggest that language‑model memorization explains the high in‑distribution success, while genuine reasoning from spectra requires powerful generation tools.